Shopify
Stripe
Slack
Notion
Google Workspace
Confluence
AI Customer Support RFP Template (2026): 46 Questions to Ask Vendors and Partners
An AI customer support RFP should give every bidder the same facts about your queue (ticket categories, volumes, help desk, channels, languages, the systems the agent must touch) and ask the same 46 questions in nine groups: product fit, accuracy and safety, actions, implementation, post-launch tuning, pricing, contract terms, security, and exit. Score the answers on a weighted sheet with a few pass/fail knockouts, then confirm the winner with a 30-day pilot on real tickets. This page gives you every part, a copyable RFP at the end, and what a good answer looks like for each question.
Key takeaways
- An AI customer support RFP should ask 46 questions in nine groups, and the pricing group must make every bidder define its billable unit in writing, including what happens on a reopen.
- Zendesk bills only resolutions an LLM verifies when an email conversation ends 72 hours after the first email, while Intercom bills $0.99 per Fin outcome, including procedure handoffs.
- Ada's customer terms, updated August 5, 2026, renew orders for one-year periods unless you give 30 days' notice, and renewal fees go up 10% for the same services.
- Post-launch help is where bids differ most: Faye's published Intercom jumpstart includes up to 12 hours over 5 business days, while a $1,700 Upwork package excludes ongoing optimization after handoff.
- Score answers on a weighted sheet with four knockouts: a signed DPA, testing on your own past tickets, a draft-only launch mode, and an export of your configuration on exit.
Disclosure: Macha publishes this blog and sells an AI agent with done-for-you setup, so we answer RFPs like this one. The template is written for any bidder: an AI vendor, a help desk partner, an agency or a freelancer. We checked every outside source on October 14, 2026.
Should you run an RFP or a pilot first?
An RFP compares bidders on paper. A pilot shows you one of them working on your tickets. Most teams need both, in that order, but the size of the deal decides how much paper is worth writing.
| Your situation | Do this | Why |
|---|---|---|
| Under about 2,000 tickets a month, one help desk, no procurement team | A short RFP (sections 1 to 4 and the pricing questions) to two or three bidders, then a pilot | A full RFP takes longer to write than a trial takes to run |
| A contract that needs a security review, a DPA and a purchase order | The full RFP, then a pilot with the finalist | Legal and IT need written answers anyway |
| An annual contract or a platform fee over five figures | The full RFP, then a paid pilot with a written exit | The auto-renewal and exit terms cost more than any pilot |
| You're also changing help desks | Run the help desk project first; our Zendesk partner guide covers that RFP | The AI bidders can't scope a desk that doesn't exist yet |
| You don't know which ticket categories an agent could take | A ticket analysis before the RFP | Bidders scope to your categories; without them, every quote is a guess |
Free trials make the "pilot first" route cheap on some products. Front's Autopilot trial covers up to 1,000 resolutions over 30 days, Fini's pricing page offers a 90-day free pilot on its Enterprise plan (both checked October 14, 2026), and Macha's trial includes the same done-for-you setup as a paid plan. A trial isn't a pilot unless someone defines success before it starts, which is what the clause near the end of this page does.
What goes in an AI customer support RFP?
Ten sections. The first six are facts you give bidders; the last four are what you ask them to send back. The more of the first six you fill in, the more comparable the answers get.
| Section | What to include | Why bidders need it |
|---|---|---|
| 1. Company context | What you sell, who your customers are, team size, hours of coverage | Tone, escalation and out-of-hours behavior depend on it |
| 2. Scope by ticket category | Your top 5 to 10 categories with monthly volume each, marked "automate", "draft for a person" or "keep human" | Bidders quote per category; "automate support" can't be priced |
| 3. Volumes | Monthly tickets by channel, peak month (Black Friday, renewals), expected growth | Usage pricing is set by these numbers |
| 4. Help desk, channels and languages | Desk and plan, channels in scope (email, chat, messaging, social, voice), languages by share of volume | Decides which products can bid at all |
| 5. Integrations and actions | Each system the agent must read or write (Shopify, Stripe, a subscription app, an internal API) and the actions with limits ("refund up to $50, once per customer") | Actions are where projects slip and where pricing changes |
| 6. Security and compliance | Your questionnaire, DPA requirement, data residency, regulated data in tickets | Starts legal review on day one instead of week six |
| 7. Pricing format requested | Setup fee, recurring fee, the billable unit with a written definition, minimums, cost at your volume and at double | Without one format, you can't compare a project quote with a monthly price |
| 8. Implementation and post-launch support | Phases with days, named people, what you must provide, hours of your team's time, who tunes the agent after launch and at what price | The ongoing work is the part most often missing from quotes |
| 9. References | Two customers whose agent has been live for 3+ months, on a similar desk | Directory reviews often predate AI work |
| 10. Timeline | Questions due date, answers due date, decision date, pilot window, target go-live | Bidders who can't meet your dates tell you something early |
Attach 30 to 50 redacted sample tickets per category if you can. Bidders who look at them answer differently from bidders who don't, and that difference is itself worth scoring.
Which questions should you ask, and what does a good answer look like?
Forty-six questions in nine groups. Send all of them to vendors. If you haven't decided which kind of provider to invite, our guide to who can set up AI customer support compares freelancers, agencies, help desk partners and vendors. For a partner or agency configuring someone else's AI, send them too, but ask the partner to answer groups 2, 6 and 8 for the AI product it would use, since that product's meter and data terms apply on top of the partner's fee.
1. Product fit (5 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 1 | Where does the agent run: inside our help desk on the ticket, or in a separate inbox or widget? | Names the desk, the channels and where replies appear; says what changes for our agents |
| 2 | Which of our categories in section 2 would you automate first, and which would you leave with people? | Picks from our list, gives a reason per category, and declines some |
| 3 | Which of our channels and languages do you support today, and which are on a roadmap? | Separates live from planned, with dates only for live features |
| 4 | Does it replace any tool we already pay for, or run alongside it? | Says what to switch off, such as overlapping help desk AI or rules, to avoid double replies |
| 5 | What is a team like ours not a good fit for? | A real answer. "We fit everyone" isn't one |
2. Accuracy and safety (6 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 6 | How do you test the agent on our past tickets before customers see it? | A test set from our closed tickets covering normal cases, escalations, edge cases and tone, with every output read by a person (method in our testing guide) |
| 7 | Can we launch in draft-only mode, with replies as internal notes until we approve? | Yes, per category, with a written bar for moving to public replies |
| 8 | What does the agent do when it doesn't know or a tool fails? | Hands off to a named group with a summary note; never guesses a policy |
| 9 | How do you stop it from taking an action we didn't allow? | Limits are written into the agent's instructions and tool permissions, per action, with amounts |
| 10 | What resolution rate do you expect on our tickets, and how will you measure it? | A range given only after reading our sample, plus the measurement method. A guaranteed percentage before seeing data is a red flag |
| 11 | How do you grade replies after launch? | Written criteria (resolution, grounding, policy, escalation, tone) applied to real conversations; see our evaluation checklist |
3. Actions and integrations (5 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 12 | Which of our systems in section 5 do you connect to today, and how (built-in connector, API, or something you build)? | One line per system, with the auth method (OAuth or a service account) |
| 13 | Which writes can the agent make, and which only read? | A list per system: "reads orders, refunds full orders with approval" beats "integrates with Shopify" |
| 14 | Who builds connections we need that you don't have, and is that in the price? | A named owner and either "included" or a price per integration |
| 15 | Whose accounts and API keys does the agent run on? | Ours, so it keeps working if the builder leaves |
| 16 | What breaks when our help desk or a connected tool changes its API or plan? | Who monitors it, and how fast a fix ships |
4. Implementation and who does the work (6 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 17 | What are the phases, with days, from kickoff to the first live category? | Named phases with day ranges and the dependencies on us (compare with our implementation timeline) |
| 18 | Who exactly does the work? | Names and roles of the people on our project, not the sales team |
| 19 | What do you need from us, and how many hours of our team's time? | Access, policies, sample tickets and a point person, with hours per week |
| 20 | Who writes the knowledge the agent answers from, and where does it live? | Gaps found from our tickets get filled before launch; articles live in our help center or our accounts |
| 21 | Who writes the agent's instructions, and can we read and edit them? | We can read every instruction; the bidder explains how edits are made |
| 22 | What's outside your scope? | A written list. Migrations, help desk redesign and custom apps are common exclusions, and that's fine if it's stated |
5. Post-launch tuning (5 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 23 | Who reads conversations after launch, how often, and for how long? | A named person or team, a weekly rhythm, and no end date, or an end date with a price after it |
| 24 | How do we request a change when a policy changes, and how fast is it live? | One channel, one person, and a turnaround in days |
| 25 | How many hours of post-launch support are included, and what does more cost? | A number and a price. For contrast, Faye's published Intercom jumpstart includes up to 12 hours over 5 business days (checked October 14, 2026) |
| 26 | Who adds the next ticket category, and is that included? | Included, or priced per category |
| 27 | What do you report to us each month? | Volume handled, escalations, misses by cause and cost, per category |
6. Pricing and billing units (6 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 28 | What is your billable unit, defined in writing? | One sentence that settles edge cases: what a conversation, resolution, outcome or ticket is |
| 29 | When does a unit count, and what happens if the customer comes back? | A time window and a reopen rule, in writing |
| 30 | Do handoffs, escalations or abandoned conversations bill? | Yes or no for each |
| 31 | What will we pay in year one at our volume and at double, all fees included? | One total: setup, platform fee, usage, seats, required onboarding, the help desk's own AI fees |
| 32 | Are there minimums, platform fees or prepaid commitments, and what happens at the limit? | Each stated with an amount; at the limit, a pause or a price, not a surprise invoice |
| 33 | Does the price change with the AI model, tool calls or conversation length? | A clear yes or no |
The definitions in questions 28 to 30 matter because vendors count differently. Since May 18, 2026, Zendesk's resolution tiers bill only a verified resolution: when the conversation ends, an LLM reads it and confirms the customer's request was "satisfactorily resolved" without escalation to a human agent. An email conversation ends 72 hours after the first email; messaging defaults to 2 hours, extendable to 72. Assisted escalations and contained resolutions don't count against the allowance. Intercom bills $0.99 per Fin outcome, and outcomes include procedure handoffs as well as resolutions, at most one per conversation, with a resolution deducted if the customer returns to the same conversation. Both definitions are defensible, and each one shapes the bill. A per-resolution meter earns the vendor more when the agent closes more, so its definition of "closed" deserves the closest reading. Our cost per resolution comparison converts the units, and our list of hidden costs and minimums covers the fees that sit outside them.
7. Contract terms (4 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 34 | What's the term, and does it renew automatically? | Monthly, or an annual term with no auto-renewal, or one with a notice window you can live with |
| 35 | What notice do we give to cancel or not renew? | A number of days, stated in the order form, not only in linked terms |
| 36 | Can the price go up at renewal, and by how much? | A cap in writing, or a fixed renewal price |
| 37 | Can we start smaller (one category, one channel) and expand without re-signing? | Yes, with the price of each step known |
Read the linked terms, not just the order form. Ada's customer terms, updated August 5, 2026, renew orders for successive one-year periods unless either side gives at least 30 days' notice, and "Fees for any Renewal Term will be subject to a 10% increase for the same Services." That's a normal enterprise clause, and the uplift pays off for the vendor whenever a customer misses the notice date. Negotiate the uplift out, or put the date in a calendar on the day you sign. Our Sierra, Decagon, Fin and Ada comparison covers those vendors' contracts.
8. Security, DPA, data retention and model training (6 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 38 | Will you sign a data processing agreement, and can we see it now? | A published or attached DPA. Under GDPR Article 28, processing on your behalf needs a binding contract |
| 39 | Which sub-processors see our ticket data, including model providers? | A named list, with where each processes data |
| 40 | Are our tickets used to train or fine-tune any model, yours or a provider's? | A written answer that covers both, and covers aggregated data too |
| 41 | How long do you keep our data, and how is it deleted at the end? | A number of days and a deletion method |
| 42 | How does the agent tell customers they're talking to an AI? | Shown wording you can approve; Article 50 of the EU AI Act sets disclosure duties from August 2, 2026 |
| 43 | Which security reports can you share (SOC 2, ISO 27001, a recent penetration test), and in which regions is our data stored and processed? | A current report under NDA, or a dated plan if one is in progress, plus named hosting and processing regions |
Specific answers to questions 40 and 41 look like Ada's terms: its third-party model providers may not "train or fine tune any such Model on Customer Data," while Ada itself "may use Aggregated and Statistical Information to improve the Models." After termination it keeps data "for a reasonable period of time" and then deletes it. Ask every bidder for that level of precision, and ask for a number where the text says "reasonable."
9. Exit and data portability (3 questions)
| # | Question | A good answer looks like |
|---|---|---|
| 44 | If we leave, what do we take with us? | Instructions, knowledge articles, tool definitions and conversation history, exportable |
| 45 | What happens to the agent and its knowledge on the day the contract ends? | A dated sequence: access removed, data exported, data deleted |
| 46 | Who owns the configuration a partner builds? | We do, documented. Smart Instinct gives that answer, "You do, documented," on its own Gorgias page |
How should you score the answers?
Copy this sheet into a spreadsheet. Score each question 0 to 3 (0 no answer or a dodge, 1 vague, 2 specific, 3 specific and in writing in the contract). Average the questions in each group, divide by 3, and multiply by the group's weight. The maximum is 100. Any knockout marked "no" removes the bid, whatever its score.
| Group | Questions | Weight | Proposal A avg | Proposal A points | Proposal B avg | Proposal B points |
|---|---|---|---|---|---|---|
| 1. Product fit | 1 to 5 | 15 | 3 | 15.0 | 2 | 10.0 |
| 2. Accuracy and safety | 6 to 11 | 20 | 2 | 13.3 | 2 | 13.3 |
| 3. Actions and integrations | 12 to 16 | 10 | 3 | 10.0 | 2 | 6.7 |
| 4. Implementation | 17 to 22 | 10 | 2 | 6.7 | 3 | 10.0 |
| 5. Post-launch tuning | 23 to 27 | 15 | 1 | 5.0 | 3 | 15.0 |
| 6. Pricing and billing units | 28 to 33 | 10 | 2 | 6.7 | 3 | 10.0 |
| 7. Contract terms | 34 to 37 | 5 | 1 | 1.7 | 3 | 5.0 |
| 8. Security and data | 38 to 43 | 10 | 3 | 10.0 | 2 | 6.7 |
| 9. Exit and portability | 44 to 46 | 5 | 2 | 3.3 | 3 | 5.0 |
| Total | 46 | 100 | 71.7 | 81.7 |
| Knockout (yes/no) | Proposal A | Proposal B |
|---|---|---|
| Signs a DPA and names sub-processors | Yes | Yes |
| Tests on our own past tickets before go-live | Yes | Yes |
| Can launch draft-only, per category | Yes | Yes |
| Configuration and knowledge exportable on exit | Yes | Yes |
In the example, A is the stronger product on paper: better fit, more connectors, a tighter security pack. B wins by 10 points because it named who tunes the agent after launch, what that costs, and a contract you can leave. We weighted post-launch tuning at 15 because that's the work that never ends, and project pricing pays for the build, so it's the part a bidder has least reason to put in a quote. If your team has a full-time admin who wants to tune the agent, move 5 points from group 5 to group 3.
As a rough guide: 80 or more is a bid you can sign after a pilot, 65 to 79 needs written answers on the low groups first, and under 65 rarely closes the gaps in negotiation. Score each proposal alone before the group discusses them, then average the scorers. This sheet scores the AI bid as a whole. For choosing a Zendesk partner for a help desk build, use the seven-row scorecard and statement-of-work checklist in our Zendesk partner guide instead.
Which answers are red flags?
- A resolution percentage before they've seen your tickets. Ask how they'd measure it instead. Smart Instinct and AY Automate both warn buyers about this on their own pages (checked October 14, 2026).
- "Unlimited" with no fair-use line. Something caps it. Ask what.
- A billable unit defined only on a sales call. If it isn't in the order form, the invoice decides.
- Post-launch support with no hours or end date. "Hypercare" needs a number. The $1,700 Upwork Gorgias package we read on October 14, 2026 lists "ongoing optimization after handoff" as not included, which is honest; vagueness isn't.
- Configuration in the bidder's accounts. If the instructions, knowledge or API keys live with them, leaving means rebuilding.
- "Integrates with" with no list of writes. Reading an order and refunding one are different projects.
- No draft-only mode. The first wrong answer goes straight to a customer.
- An auto-renewal with a renewal uplift and a short notice window, buried in linked terms.
- The people in the pitch aren't the people on the project. Ask for names in the statement of work.
- No references live for three months or more. A two-week-old agent hasn't met a policy change yet.
How do you add a pilot clause?
Put this in the RFP so every bidder prices the pilot the same way, then attach it to the finalist's order form. Our 30-day pilot plan shows how to run it day by day, with a go/no-go scorecard. Our rollback post covers why pilots get switched off and what to measure so yours isn't.
PILOT / PROOF OF CONCEPT
1. Scope: one ticket category ([name]), one channel, [n] tickets or 30 days, whichever comes first.
2. Mode: replies are drafted as internal notes for our team to approve. Public replies only after
we confirm in writing that the bar in (4) is met.
3. Before go-live: the bidder tests on [50-100] of our past tickets in that category and shares
every output.
4. Success bar (agreed before start): [e.g. 9 in 10 drafts sent unchanged or with light edits over
a full week]; zero actions outside the written limits; handoffs reach the right group with a note.
5. Cost: [free / fixed fee], credited in full against the first invoice if we sign.
6. Exit: either side may end the pilot with 2 business days' notice. On exit the bidder exports
the instructions and knowledge built for us and deletes our data within [x] days.
7. Weekly: a 30-minute review of misses grouped by cause.Copy the full RFP
Paste this into a doc. Fill in sections 1 to 6; bidders answer 7 to 10 and the 46 questions.
AI CUSTOMER SUPPORT RFP (template: getmacha.com, Oct 2026)
Issued by: [company] Contact: [name, email]
Questions due: [date] Answers due: [date] Decision: [date] Pilot: [dates] Go-live: [date]
1. COMPANY CONTEXT: what we sell, customers, team size, support hours
2. SCOPE BY CATEGORY: top 5-10 categories, monthly volume each,
marked automate / draft for a person / keep human. Sample tickets attached.
3. VOLUMES: monthly tickets by channel, peak month, expected growth
4. HELP DESK, CHANNELS, LANGUAGES: desk and plan; channels in scope; languages by share
5. INTEGRATIONS AND ACTIONS: systems to read/write; each action with its limit
6. SECURITY AND COMPLIANCE: questionnaire attached; DPA required; residency; regulated data
7. PRICING FORMAT (bidder): setup fee; recurring fee; billable unit defined in writing;
minimums; year-one total at our volume and at double, all fees included
8. IMPLEMENTATION AND POST-LAUNCH (bidder): phases with days; named people; our hours;
who tunes after launch, for how long, at what price
9. REFERENCES (bidder): two customers live 3+ months on a similar help desk
10. PILOT: see attached pilot clause
QUESTIONS
Product fit
1 Where does the agent run: in our help desk on the ticket, or a separate inbox/widget?
2 Which of our categories would you automate first, and which leave with people?
3 Which of our channels and languages are live today, and which are roadmap?
4 What does it replace or run alongside? What should we switch off?
5 What is a team like ours not a good fit for?
Accuracy and safety
6 How do you test on our past tickets before customers see anything?
7 Can we launch draft-only (internal notes) per category?
8 What happens when the agent doesn't know or a tool fails?
9 How do you stop actions we didn't allow?
10 What resolution rate do you expect on our tickets, and how will you measure it?
11 How do you grade replies after launch?
Actions and integrations
12 Which of our systems do you connect to today, and how?
13 Which writes can the agent make, and which only read?
14 Who builds missing connections, and is that in the price?
15 Whose accounts and API keys does the agent run on?
16 What breaks when our desk or a tool changes its API or plan, and who fixes it?
Implementation
17 Phases with days from kickoff to the first live category?
18 Who exactly does the work (names, roles)?
19 What do you need from us, in hours per week?
20 Who writes the knowledge, and where does it live?
21 Who writes the agent's instructions, and can we read and edit them?
22 What's outside your scope?
Post-launch tuning
23 Who reads conversations after launch, how often, for how long?
24 How do we request a change, and how fast is it live?
25 Post-launch hours included, and the price after?
26 Who adds the next category, and is it included?
27 What do you report monthly?
Pricing and billing units
28 Your billable unit, defined in writing?
29 When does a unit count, and what happens on a reopen?
30 Do handoffs, escalations or abandoned conversations bill?
31 Year-one total at our volume and at double, all fees included?
32 Minimums, platform fees, prepaid commitments; what happens at the limit?
33 Does price change with the model, tool calls or conversation length?
Contract terms
34 Term length; automatic renewal?
35 Notice to cancel or not renew?
36 Renewal price increase, and its cap?
37 Can we start with one category/channel and expand without re-signing?
Security, DPA, retention, training
38 Will you sign a DPA? Attach it.
39 Sub-processors seeing ticket data, including model providers, and where?
40 Are our tickets used to train or fine-tune any model, yours or a provider's?
41 Retention period and deletion method at contract end?
42 How does the agent disclose it's an AI? Show the wording.
43 SOC 2 / ISO 27001 / pen test reports? Data storage and processing regions?
Exit and portability
44 What can we export if we leave?
45 What happens to the agent and data on the last day, step by step?
46 Who owns configuration a partner builds?
SCORING: 0-3 per question; group average / 3 x weight (fit 15, accuracy 20, actions 10,
implementation 10, tuning 15, pricing 10, contract 5, security 10, exit 5).
Knockouts: DPA; testing on our past tickets; draft-only launch; export on exit.Which parts matter most for your situation?
| You are | Weight up | Ask hardest |
|---|---|---|
| A Shopify brand on Gorgias, no developer | Actions (15), tuning (20) | 13 (which writes), 24 (policy-change turnaround), 32 (what happens in your peak month) |
| A B2B SaaS team on Intercom comparing Fin with others | Pricing (15) | 28 to 30 on units, and 4 on what to switch off |
| An EU team with a procurement process | Security (20) | 38 to 43, with the DPA attached before the shortlist |
| An enterprise signing an annual contract | Contract (10), exit (10) | 34 to 36 and 44 to 46; get the renewal cap in the order form |
| A small team, one admin, under 1,000 tickets a month | Tuning (20), implementation (15) | 19 (your hours) and 23 (who reads conversations); skip the full RFP and run a pilot |
| A team hiring a partner to configure the help desk's own AI | Implementation (15), tuning (20) | 22 (scope), 25 (hours after launch), 46 (ownership), plus groups 6 and 8 for the AI product's own terms |
Keep each row's total at 100 by taking the extra points from the groups that row doesn't name.
Where Macha fits
Macha is ours, and it's one of the bidders this template is written for. It runs on Zendesk, Freshdesk, Gorgias, Front, HubSpot or Intercom. Short answers to the questions where bids differ most: the agent works inside your help desk on the ticket (1); the Macha team tests on your past tickets and launches each category in safe mode, with drafts as internal notes until your team says they're right (6, 7); done-for-you setup covers the ticket analysis, knowledge base, agent and custom API tools after about 15 minutes of you connecting your desk and tools (14, 19); and a dedicated success manager handles policy changes, new categories and tuning after launch at no extra charge (23 to 26).
On pricing (28 to 33), the unit is the ticket: one thread with one person, one charge however many messages it takes, whether or not the agent resolves it, and the model doesn't change the price. That's about $0.40 per ticket, from $299/month for 750 tickets, billed monthly with no setup fee (pricing); at the limit, agents pause rather than bill overages. Our DPA is published (38), and SOC 2 is in progress, expected December 2026 (43). The promise is "live in safe mode within 7 days, or your first month is free". The trial includes the same setup, with $50 of free usage (about 125 tickets), no credit card, no time limit, which makes it a reasonable way to run the pilot clause above. It suits teams already on one of those desks who want the agent built and tuned by someone else; it's the wrong fit if you need a help desk migration or a desk we don't run on. To get section 2 of this RFP filled in from your own tickets, ask for the free support audit.
How we researched this
The ten sections and 46 questions come from the questions buyers ask in the implementation pages we studied on October 14, 2026 (agency, vendor and marketplace pages for Zendesk, Gorgias, Intercom, Freshdesk, HubSpot and Front), from our live implementation guide, checklist and cost breakdown. On October 14, 2026 we read Zendesk's automated resolution tiers article, Intercom's Fin outcomes article, Ada's customer terms of use (updated August 5, 2026), Article 28 of GDPR and Article 50 of the EU AI Act. The Faye, Upwork and Smart Instinct figures come from their own pages, read the same day for our partner ranking; the Front and Fini trial terms from their own pricing pages, also October 14. The "good answer" column, the scoring weights, the knockouts, the worked example and the pilot clause are our recommendations, not a published standard. This page isn't legal advice; have counsel review the contract and data terms.
Common questions
What should an AI customer support RFP include?
Six sections you fill in (company context, scope by ticket category, volumes, help desk with channels and languages, integrations and actions, security requirements) and four the bidder answers (pricing in a set format, implementation and post-launch support, references, timeline), plus a question list and a scoring sheet. The category list with volumes does the most work, because it's what bidders price against.
How many questions should an AI vendor RFP have?
Enough to cover nine areas without padding: product fit, accuracy and safety, actions, implementation, post-launch tuning, pricing, contract, security and exit. This template uses 46. A small team can send groups 2, 5 and 6 alone, about 17 questions, and learn most of what matters.
What's the most important question to ask an AI customer support vendor?
Who changes the agent after launch, how fast, and whether that's in the price. Every bidder can describe a build; fewer will put a price on the work that continues for as long as the agent runs.
How do you compare vendors that bill per resolution, per conversation and per ticket?
Ask each for a written definition of its unit and a year-one total at your volume and at double, with every fee included. Then convert to cost per resolved ticket at the same assumed resolution rate. Zendesk and Intercom define their units differently enough that the list prices alone don't compare.
Should we run an RFP or a pilot?
Both, in that order, if the contract needs a security review or runs a year. Under about 2,000 tickets a month with monthly billing, a short RFP to two or three bidders and a 30-day pilot with success criteria agreed up front is faster and tells you more.
Can we send the same RFP to an AI vendor and an implementation partner?
Yes, and you should if the AI agent is the whole project. Ask the partner to answer the pricing, accuracy and data questions for the AI product it would configure, since that product's meter and terms run on top of the partner's fee.
