Picture a service director at a mid-size retailer who receives 3 Agentforce proposals in the same week. The first is a $60,000 fixed-price pilot, the second is time and materials with no stated total and the third is a per-conversation arrangement that never mentions testing. All 3 firms call themselves Agentforce experts, all 3 offer a reference and none says which metric will count as success. The director has 2 weeks to pick a finalist and no way to tell which proposal carries the most risk. This scenario is invented for illustration and doesn’t describe a real company.

This post gives 12 questions to ask an Agentforce partner, with what a good answer sounds like and where each answer can mislead. It draws on Salesforce’s Agentforce pricing page (modified Oct 8, 2026), its partner program launch document and partner page, a Trailhead unit on the Einstein Trust Layer, a trade-press report on a Gartner release, Fortune’s summary of an MIT report and guides from Argano, Mavlers, Traction Complete, Atlan and Apex Hours. Most of those guides come from firms selling services or software, so I flag the bias as I go. I found no neutral data comparing partner quality, so every scoring weight and budget figure in this post is invented. Nothing in this post is legal advice, so confirm contract terms with counsel and prices with your Salesforce account executive.

TL;DR

  • Ask for proof before credentials. Salesforce’s partner tiers (Base, Ridge, Crest and Summit) are earned through a Trailblazer score, and the document I read doesn’t say what each tier requires. A badge narrows a list, but a reference call with a customer running an agent in production tells you far more about delivery.
  • Make the partner price the consumption as well as the build. Salesforce lists an action at 20 Flex Credits ($0.10) and a conversation at $2, so a single agent handling 30,000 conversations a month can cost $24,000 or $60,000, depending on how many actions it takes. A quote that ignores this hides the larger bill.
  • Score the answers with weights you set first. In my invented sheet, 2 partners tie at 42 raw points out of 60 and separate only after weighting, at 69 versus 70 out of 100. That gap is a reason to keep asking questions rather than a verdict.

What the sources say, and who published them

The table lists the sources behind this post, who published each and how far I’d trust it. Salesforce is the only primary source for prices and program rules, and it also sells the product, so its pages describe what it wants buyers to believe about partners. Trade press and consultancies mostly repeat other people’s numbers, and firms that sell implementation have an obvious reason to describe failure as something their own service prevents.

SourcePublisher typeWhat I used it forBias or limit
Agentforce pricing page (modified Oct 8, 2026)Vendor, primaryCredits, add-ons, per-conversation priceSays pricing is subject to change
Partner program launch documentVendor, primary, undatedTier names, scorecard pillars, distinctionsDoesn’t explain what each tier requires
Partner pageVendor, primaryPartner types, partner finder, quoted resultsQuoted results are marketing
Trailhead Trust Layer unitVendor trainingMasking, retention, audit trailSalesforce describing its own product
w.media report on Gartner (June 27, 2025)Trade press, secondary40 percent cancellation forecast, agent washingI didn’t read Gartner’s release
Fortune on MIT (Aug 18, 2025)Press, secondary5 percent, 67 percent findingsReport not read, failure undefined
Argano (Feb 23, 2026)Services firmPitfalls listSells Agentforce services
Mavlers (Feb 24, 2026)Services firm5 failure causesSells services, several claims unsourced
Traction Complete (updated Sept 8, 2026)Data software vendorData readiness checksStatistics from its own surveys
Atlan (Aug 4, 2026)Data governance vendorGovernance gapsProduct pitch
Apex Hours (Feb 25, 2025)Independent training blogTesting CenterLeans on a Salesforce press release

The pattern is plain. Nobody neutral has measured how well Agentforce partners deliver, so the evidence on partner quality is anecdote, vendor marketing and general studies of AI projects. That’s why the 12 questions focus on things you can check yourself: references you can call, documents you can read, tests you can watch run and prices you can recompute. Verify each claim from a firm that profits from the answer before you rely on it.

Why the partner choice carries so much weight

Gartner’s June 2025 release, as a trade outlet reported it, says more than 40 percent of agentic AI projects could be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. The poll behind it covered 3,412 webinar attendees in January 2025, of whom 19 percent reported significant investment, 42 percent conservative investment and 8 percent none, while the rest were waiting or unsure. That’s a self-selected sample, and a forecast isn’t a count, so I treat the 40 percent as a warning rather than a measurement.

The same report says Gartner estimates only about 130 of the thousands of agentic vendors are real, with the rest rebranding assistants, chatbots and robotic process automation, a habit it calls agent washing. The label describes software vendors, but a similar pressure plausibly applies to service firms, since an Agentforce badge costs less to claim than a production deployment costs to run. That second point is my inference rather than a Gartner finding, and it’s the reason Question 1 asks for production proof.

Fortune’s summary of an MIT report puts the share of generative AI pilots that reach rapid revenue acceleration at about 5 percent, and says tools bought from specialized vendors or built through partnerships succeeded about 67 percent of the time, roughly 3 times the rate of internal builds. The report drew on about 150 interviews, a survey of 350 employees and 300 public deployments. Fortune doesn’t define failure, I read only its summary, and the study covers generative AI pilots in general rather than Agentforce, so it supports hiring help without telling you whom to hire.

Questions 1 to 3: proof and credentials

  • Question 1: Which agents have you taken to production, and may I speak to those customers? Ask for the use case, the go-live date, the volume the agent handles and the metric the customer tracks. A partner with real deployments can name them or arrange a call, while a firm that offers only logos or an unnamed large retailer hasn’t answered. Salesforce’s partner page quotes results such as 84 percent autonomous case resolution and $15M in pipeline for named partners, but those are marketing claims with no method behind them, so use them to form questions rather than as benchmarks.
  • Question 2: Which Salesforce tier and Agentforce distinctions do you hold, and who on my team holds the certifications? Salesforce’s partner program launch document lists the tiers Base, Ridge, Crest and Summit, says tier is earned from a Trailblazer score across customer success, innovation, growth and impact, and limits Summit to consulting and AppExchange ISV partners. It also mentions Implementation Expert and Cloud Expert distinctions, including a new Agentforce distinction. The document gives no metrics or requirements and carries no date, so ask the partner to show its current listing and name the certified people assigned to your project.
  • Question 3: Who will do the work, and how many Agentforce projects has each of them delivered? Proposals are often written by senior people and delivered by others. Ask for named roles, hours per week and Agentforce experience for each, and put replacement rules in the statement of work so that a substitution needs your approval. Salesforce’s partner page also lists a separate network of forward-deployed engineering partners, so ask whether the firm builds with its own staff, subcontractors or such a partner, because the answer decides who you can hold accountable for the result.

Questions 4 to 6: scope, data and integration

Question 4: How will you choose and rank the first use case? Argano and Mavlers both advise starting small. Argano suggests a single solid use case such as FAQ automation or order inquiries, and Mavlers proposes a minimum viable agent with a single high-impact, low-complexity use case, a named business owner and a success metric. Ask the partner for its ranking criteria, such as volume, repeatability, data availability and risk, and for the business owner it expects on your side. A partner proposing to automate everything at once is describing what Argano calls overreach, and should explain how it plans to control that risk.

Question 5: What will you check in my data and knowledge before building? Argano says most knowledge bases aren’t ready for agents, Mavlers recommends flagging fields with more than 10 percent null values, a threshold it doesn’t source, and Traction Complete lists duplicates, account hierarchy, field consistency, completeness and staleness as checks. Traction Complete sells data tools and bases its statistics on its own surveys, so read them as a vendor’s view. Whatever the source, ask for a written data and knowledge assessment as a priced deliverable, with named fixes and owners, before any build begins.

Question 6: Who owns the integrations, and what happens when an upstream API changes? Argano says agents can’t complete actions across orders, shipments and pricing without MuleSoft or direct API access, and Mavlets names third-party API failures, unnoticed upstream changes and technical debt that hits governor limits under agent load. Both firms sell integration work. Even so, the questions are fair: which systems connect, who builds and monitors each connection, whether MuleSoft or direct API access is needed and priced separately, and how a change in a connected system reaches the team before customers notice.

Questions 7 to 9: testing, safeguards and measurement

Apex Hours, an independent Salesforce training blog, describes the Agentforce Testing Center as a way to build and test agents at scale, in a sandbox, with test scenarios per agent that are uploaded or generated and checked for the right topic and action sequence, with a pass or fail outcome. The article rests largely on a Salesforce press release and doesn’t discuss limits, so ask the partner to show the tool running on your own queries. The table sets out what a good answer contains and what should worry you.

QuestionA good answer includesA warning sign
7. How will you test before launch and after every change?A written test set built from your real queries, run in a sandbox, with pass or fail on topic and action sequence and a rerun after each change“We test during user acceptance” with no test set, owner or pass mark
8. How do you handle permissions, masking and audit?Role-based personas, field-level security, least-privilege access, human approval for high-stakes actions and a demo of what gets masked in your sandboxBroad access granted early, or “the Trust Layer handles it” with no demonstration
9. Which metric defines success, and how do we baseline it?A single primary metric, a baseline measured before the build, a named business owner and a review dateActivity counts such as conversations handled, with no link to cost or resolution

Salesforce’s own Trailhead unit on the Einstein Trust Layer says the Trust Layer masks sensitive data before a prompt reaches the model, restores it in the reply, relies on agreements under which providers don’t retain prompts or responses, scores toxicity and keeps an audit trail. Atlan, a data governance vendor, says the layer doesn’t mask data for agents and that Salesforce documents this. I couldn’t reconcile those accounts, so ask the partner to demonstrate what gets masked in your sandbox, and read Salesforce’s help documentation yourself before you rely on either account.

How the platform bills, and why the partner’s quote depends on it

Salesforce’s Agentforce pricing page, published Oct 30, 2024 and modified Oct 8, 2026, says it’s provided for information only and is subject to change. It lists Flex Credits at $500 per 100,000, with an Agentforce action costing 20 credits and a voice action 30. Per-user add-ons cost $125 per user per month for Sales, Service, Field Service and Operations, with unmetered employee usage, and $150 for industry add-ons. Max Editions start at $550 per user per month with 2.75 million credits per organization per year, a user license costs $5 and requires Flex Credits, and conversations cost $2 each.

The arithmetic shows why design choices set the bill. At 20 credits and $500 per 100,000, an action costs $0.10, so a conversation breaks even with the $2 price at 20 actions. For an invented 30,000 conversations a month, credits cost $12,000 at 4 actions per conversation, $24,000 at 8, $36,000 at 12, $60,000 at 20 and $75,000 at 25, against $60,000 at $2 per conversation. A partner who designs an 8-action flow and another who designs a 25-action flow produce a bill about 3 times apart for the same business result.

Seat-based costs behave differently. At $125 a month, 25 users cost $37,500 a year, 60 users cost $90,000 and 150 users cost $225,000, and 25 users on the $150 industry add-on cost $45,000. The included pool is also small next to customer-facing volume, since 2.75 million credits cover 137,500 actions a year, which is about 1,432 conversations a month at 8 actions each. I haven’t verified how Salesforce meters every edge case, so ask the partner to model each channel from your volumes and to show the assumption behind each line of the quote.

Questions 10 to 12: commercial terms and life after launch

Question 10: Is the price fixed, time and materials or tied to outcomes, and what’s excluded? A fixed fee suits a scope that discovery has defined, time and materials suits uncertain work if it carries weekly hour caps and a change-approval path, and outcome-based pricing needs a baseline both sides trust. Ask for a list of exclusions, covering licenses, credits, integration software, data cleanup and training. Get every exclusion written down before you compare totals.

Question 11: How will my administrators take over, and what will you leave behind? Argano advises planning for tuning against real user queries because perfect instructions on day 1 are unrealistic, and Mavlers recommends role-specific training and a weekly feedback triage. Both are vendor advice, but the handover point holds. Ask for documentation, the test set, the instruction history and a runbook as named deliverables, and for a defined period of supported tuning, so that your team can change the agent without calling the partner for every edit.

Question 12: What happens after launch, and how do we part ways if it isn’t working? Ask what support is included, how long it lasts, what a change request costs and who monitors the agent. Then ask about exit: who owns the configuration and documentation, what notice ends the contract and what the partner hands over on the way out. A partner confident in its work will answer plainly. I haven’t reviewed enforceability in any jurisdiction, so have counsel read these terms before you sign.

Red flags in a proposal

  • No named references, or only logos. A firm that has taken agents to production can usually arrange a call. If every example is anonymous, if results are quoted without a method or if the numbers resemble the unverified ones in marketing posts, such as Argano’s 40 percent adoption gain for an unnamed client, treat the claims as unproven. Ask for a customer in your industry or at your volume, and ask your Salesforce account executive whether a second reference exists.
  • A fixed price before any discovery. A quote that fixes the total before anyone has looked at your data, systems or volumes is either padded or hiding assumptions that will turn into change orders. Look for a paid discovery phase with deliverables you own, such as a ranked use case list, a data assessment and an integration list, so that you can take the output to another bidder if the first partner’s follow-on quote disappoints you.
  • A test plan that starts at user acceptance. If testing appears only as a final phase with your staff doing the checking, the partner hasn’t planned for regressions, and every later change will be tested by your customers. Ask to see the test set, the pass mark and who reruns the tests after each change. Make that a condition of any contract. Ask what happens to results when a model or Salesforce release changes the agent’s behavior.

What badges and partner finders can and can’t tell you

Salesforce’s partner page describes an AgentExchange ecosystem of technology, solution, data, model and consulting partners, and a “Find a Consulting Partner” link that promises certified partners matched to your needs. It also lists a separate network of forward-deployed engineering partners. The page doesn’t describe badges, tiers or specializations, and the filtering behind the finder isn’t explained. A finder is useful for building a long list by region, industry and product, but it can’t tell you who delivers well.

The practical approach is to use the finder and tiers to build a list of 4 to 5 firms, then discard the badge and test the answers. Salesforce’s page also notes that its Customer Success Group, a team of Salesforce employees, works with partners to help customers launch, so ask whether that group will be involved and what it does. Their answers matter more than their tiers, because only the answers show how the firm works.

A scoring sheet with weights you set first

The sheet scores 2 invented partners against the 12 questions on a scale of 0 to 5, with weights that sum to 100. I chose the weights to show how weighting works, not to recommend them. Set yours before you read any proposal, because weights chosen afterward tend to favor whichever partner you already liked. Each weighted cell is the weight times the score, so the maximum is 500 points, which I divide by 5 to report out of 100.

QTopicWeightA scoreB scoreA pointsB points
1Production references8534024
2Credentials7432821
3Named team6351830
4First use case9433627
5Data and knowledge checks9241836
6Integration8342432
7Testing8433224
8Safeguards10343040
9Success metric8524016
10Commercial terms9342736
11Handover8433224
12After launch and exit10242040

Total1004242345350

The raw totals tie at 42 of 60, and the weighted totals are 345 and 350 of 500, or 69 and 70 out of 100. A gap of 1 point out of 100 is a tie in practice. Partner A shines on proof and on the success metric, while Partner B is stronger on data, integration, safeguards and exit. The sheet’s value is in showing where to probe next, so run follow-up calls on the lowest scores and rescore, instead of declaring a winner from the first pass.

Running the reference call

Ask about the work rather than the relationship. Good questions cover the time from kickoff to production, what changed between pilot and launch, how many tuning cycles followed, what broke and how the partner responded. Ask what the final bill was against the quote, including credits and integration software. Ask what the customer would change in the contract. These questions pull out specifics, and people who ran a real project can answer them without preparation.

Listen for specifics. A useful reference names the metric, the baseline and the result, mentions a test set or a review routine and describes at least 1 thing that went wrong. A reference that only praises the team is telling you about rapport. Compare what you hear with what the proposal claims, since a gap between the partner’s numbers and the customer’s numbers is the most valuable thing a call can reveal, and write down both before the call ends.

Budgeting a pilot and the first year

Take an invented service team with 30,000 conversations a month, of which an agent resolves 10,000 without a human, or 33.3 percent. Assume a saving of $4.00 per resolved case, a fixed pilot fee of $60,000 and 8 actions per conversation. On credits, the agent costs 30,000 × 8 × $0.10 = $24,000 a month, because every conversation consumes credits even when a human takes over. The savings is 10,000 × $4.00 = $40,000, leaving a net of $16,000 a month and a payback on the pilot fee of 3.75 months.

Change only the billing model and the sign flips. At $2 per conversation, the same 30,000 conversations cost $60,000 a month, which exceeds the $40,000 saving by $20,000. Nothing about the agent changed, but its economics did. That’s why a proposal that shows a build price and no monthly run cost is incomplete, and why the partner should build this model from your volumes during discovery and update it at every design decision that changes actions per conversation.

Every figure above is invented, including the saving per case and the resolution rate, and real results depend on your cases, your agent design and Salesforce’s metering rules. The structure is what matters: build fee, monthly consumption, monthly saving and payback, all stated with their assumptions. Ask each finalist to produce this table for your business, compare their assumptions line by line and treat a partner who refuses to state a resolution rate as a firm that hasn’t thought about the answer.

What to put in the statement of work

  • Scope, exclusions and assumptions. List the use cases, channels, systems and data in scope, and state what’s excluded, such as licenses, credits, integration software and data cleanup. Add the assumptions behind the price, such as conversation volume and actions per conversation, and a rule that a change in any assumption triggers a written change request. Without these, the partner can describe nearly any later request as outside scope, and you can’t argue the point.
  • Acceptance criteria and the test set. Define what passing means before the build starts: the test set, the pass mark, the primary metric and the date it’s measured. Make the test set a deliverable you own, and require a rerun after every change that touches the agent. Acceptance tied to a date alone invites a launch that meets the schedule and misses the goal, and the partner has no reason to care about the difference.
  • Handover, support and exit. Specify the documentation, runbook, instruction history and training you’ll receive, the length and cost of post-launch support and what happens on termination. Confirm who owns the configuration, and ask counsel to review the terms, since I haven’t reviewed enforceability in any jurisdiction. An exit clause is easiest to negotiate before the relationship is strained, and the partner’s reaction to it tells you plenty. Get all of it in writing before work begins.

A 12-week timeline from kickoff

An invented plan shows what milestones to expect. I found no neutral data on Agentforce project durations. Argano says focused pilots can show returns in weeks rather than months and describes a 4-week program of its own, both without supporting data, and Mavlers gives no durations at all. The schedule below starts on Monday, Jan 11, 2027, and the dates are computed from that start. Treat the shape as a template and the week counts as assumptions to test with each finalist.

WeekDateMilestoneOwner
0Mon Jan 11, 2027Kickoff, sandbox access, named business ownerBoth
2Mon Jan 25, 2027Discovery signed off: first use case, metric and baselineYou and partner
4Mon Feb 8, 2027Data and knowledge assessment delivered, fixes scheduledPartner and your data owners
6Mon Feb 22, 2027First run against the written test setPartner
8Mon Mar 8, 2027Sandbox acceptance testing with your staffYou
12Mon Apr 5, 2027Limited production release to a share of trafficBoth

The data fixes in week 4 and the acceptance testing in week 8 tend to slip, because both depend on your staff’s time. Ask the partner what happens to the dates if access arrives late, and tie any license or credit purchases to the release rather than to the kickoff, so that a slip doesn’t turn into months of paying for capacity nobody is using. Write the milestones and their owners into the statement of work.

Making the decision

Narrow the field to 2 finalists after the first round of answers, then pay for a short discovery with each if the budget allows, or with the stronger finalist if it doesn’t. Discovery shows how the team works, produces a data assessment and a use case ranking you own, and gives a real basis for a fixed price. A partner unwilling to scope before quoting a total is telling you how it treats uncertainty, and you can learn that cheaply before committing to a build.

Be honest about what remains unknown. I couldn’t confirm what each Salesforce tier requires, whether the launch document is current, how Salesforce meters every usage case or what the Trust Layer masks for agents, and I found no neutral study of partner quality. Before you sign, confirm tiers and pricing with your Salesforce account executive, have counsel review the contract and ask the partner to demonstrate, on your own data in a sandbox, everything it has claimed.

FAQs

What’s the most important question to ask an Agentforce partner?

Ask which agents the firm has taken to production and whether you can speak to those customers. Proof of delivery tells you more than a badge, a logo wall or a pitch deck, because it shows how the partner handles real data, real volumes and real problems after launch. Follow up with the metric, the baseline and what went wrong. I found no neutral data comparing partners, so references are the best evidence available.

How much does an Agentforce implementation cost?

I found no neutral figure for implementation services. Salesforce’s pricing page covers licenses and credits only, listing an action at 20 Flex Credits ($0.10), a conversation at $2 and add-ons at $125 to $150 per user per month. Consultancy guides I read gave no reliable service ranges. Ask each partner for a priced discovery and a model of monthly consumption, and compare quotes with licenses and credits included. Treat any range from a seller as marketing.

Should I pick the partner with the highest Salesforce tier?

Not automatically. Salesforce’s launch document says tier is earned from a Trailblazer score across customer success, innovation, growth and impact, and limits Summit to consulting and AppExchange ISV partners, but it doesn’t say what each tier requires or whether it predicts your result. Use tier to build a list, then test the answers: references, named staff, a test plan and a priced model of consumption. Tier is a starting point for a list, and your own checks decide the shortlist.

How many partners should I interview?

I’d talk to 4 or 5 firms, ask all of them the same 12 questions in writing and then select 2 finalists. That’s my judgment rather than a researched figure. More than 5 conversations tend to produce comparison fatigue, while fewer than 3 leave you without a baseline for price and approach. Set the scoring weights before the first call so that you judge every firm by the same standard.

Do I need a partner, or can my administrators build it?

It depends on your team’s experience, the use case and the integrations. Fortune’s summary of an MIT report says tools bought from specialized vendors or built through partnerships succeeded about 67 percent of the time, roughly 3 times the rate of internal builds. I read only the summary, the study covers generative AI pilots in general and failure isn’t defined, so treat it as a signal rather than proof. Ask a partner to explain the trade-off for your case.

What’s agent washing?

It’s Gartner’s term for vendors rebranding existing assistants, chatbots or robotic process automation as agentic AI without real agentic capability. A trade report on Gartner’s June 2025 release says only about 130 of the thousands of agentic vendors are real. The warning concerns software vendors, though service firms could do something similar with badges and slogans. Ask for production proof, a working demo on your data and named references. Ask any partner to show a running agent on your data.

How do I compare pricing models?

Convert each quote to a monthly run cost at your own volume. At 20 credits and $500 per 100,000 credits, an action costs $0.10, so a conversation on credits costs the same as Salesforce’s $2 conversation price at 20 actions. For 30,000 conversations a month, 8 actions cost $24,000 on credits against $60,000 per conversation. Add per-user licenses, then confirm all rates with your account executive, since the page is subject to change.

What should the contract include?

At minimum: scope and exclusions, the assumptions behind the price, acceptance criteria with a test set you own, named team members with substitution approval, handover deliverables, a defined support period and exit terms covering ownership of configuration and documentation. Put the monthly consumption model in an appendix. I’m not a lawyer and haven’t reviewed enforceability in any jurisdiction, so have counsel read the final agreement. Keep a copy of the signed assumptions with the order form.

By

Leave a Reply

Your email address will not be published. Required fields are marked *