All industriesAI & operations
How to choose an AI implementation partner (questions to ask before you sign)
By Yath Regmi, Founder, Kodec
Verified on Updated Published
Short answer
The right AI implementation partner will name the finished job in hours saved and dollars kept, hand you the system, and say who runs it after launch. A partner badge shows a firm joined a program, which says little about whether it can do your job. Large firms are reported at $80K to $300K and up. A smaller shop fits one process first. Score every candidate on the same 12 questions below.
When this is the right call
A big firm is the better choice in some cases. If you run a company with many sites, strict audit rules, or an AI program that spans finance, legal and operations, pick Accenture, Deloitte, BCG X, IBM, Slalom or a firm like them. They bring large teams, security reviews and programs that last years. If you want one process fixed first, with a price and a date, a smaller firm usually fits better. Ask for the same proof from both.
What an AI implementation partner should deliver, in hours and dollars
Example
| Line | Example |
|---|---|
| Staff hours saved per week | 20 hours |
| Loaded cost per hour | $40 |
| Value per year (20 x $40 x 52) | $41,600 |
| One-time build cost | $30,000 |
| Run cost per month (hosting, model fees, support) | $500 |
| Run cost per year | $6,000 |
| Net saving per month (41,600 / 12 - 500) | $2,967 |
| Months to pay back the build (30,000 / 2,967) | about 10 |
| Year-one value per $1 spent (41,600 / 36,000) | $1.16 |
All inputs are invented. Swap in your own hours and pay. Ask each partner to fill in this table before you sign, then measure the same lines 90 days after launch. A demo that works on stage is not on this list.
Side by side
OpenAI
| Program and levels | OpenAI Partner Network, announced June 14, 2026. Select, Advanced, Elite.4 |
|---|---|
| What a firm must show | Sales, technical capability, co-sell with OpenAI and deployment experience. No numeric thresholds are published in the source.4 |
| What the badge does not tell you | The source calls certification necessary but not sufficient. A badge does not define delivery quality.4 |
Anthropic
| Program and levels | Claude Partner Network, launched March 12, 2026. Select, Preferred, Global Premier.2 |
|---|---|
| What a firm must show | Select needs 10 certified people, 2 customers in production and 1 public story. Preferred needs 100, 15 and 3. Global Premier needs 1,000, 100 and 15.1 |
| What the badge does not tell you | Anthropic says the tier measures built practice, not outcomes.1 |
AWS
| Program and levels | Partner tiers Select, Advanced, Premier, plus a Generative AI Competency.6 |
|---|---|
| What a firm must show | For the Competency, AWS validates business and technical expertise in generative AI.6 |
| What the badge does not tell you | It is a firm-level check, not a promise about your project.6 |
Microsoft
| Program and levels | Solutions Partner designations and specializations.7 |
|---|---|
| What a firm must show | Points for deal performance, certifications and customer success. Thresholds are not in the source.7 |
| What the badge does not tell you | A designation shows points earned. It does not name which of the firm's people will build your system.7 |
Google Cloud
| Program and levels | Select, Premier, Diamond.7 |
|---|---|
| What a firm must show | Customer success and outcomes, plus certifications and closed deals. Weights are not documented in the source.7 |
| What the badge does not tell you | Same limit. A tier is a program position, not a result for you.7 |
| What it is | OpenAI | Anthropic | AWS | Microsoft | Google Cloud |
|---|---|---|---|---|---|
| Program and levels | OpenAI Partner Network, announced June 14, 2026. Select, Advanced, Elite.4 | Claude Partner Network, launched March 12, 2026. Select, Preferred, Global Premier.2 | Partner tiers Select, Advanced, Premier, plus a Generative AI Competency.6 | Solutions Partner designations and specializations.7 | Select, Premier, Diamond.7 |
| What a firm must show | Sales, technical capability, co-sell with OpenAI and deployment experience. No numeric thresholds are published in the source.4 | Select needs 10 certified people, 2 customers in production and 1 public story. Preferred needs 100, 15 and 3. Global Premier needs 1,000, 100 and 15.1 | For the Competency, AWS validates business and technical expertise in generative AI.6 | Points for deal performance, certifications and customer success. Thresholds are not in the source.7 | Customer success and outcomes, plus certifications and closed deals. Weights are not documented in the source.7 |
| What the badge does not tell you | The source calls certification necessary but not sufficient. A badge does not define delivery quality.4 | Anthropic says the tier measures built practice, not outcomes.1 | It is a firm-level check, not a promise about your project.6 | A designation shows points earned. It does not name which of the firm's people will build your system.7 | Same limit. A tier is a program position, not a result for you.7 |
Vendor by vendor
Fixed-price pilot
Layer3Labs says its fixed-price pilots start at $10,000 and lists a single-workflow pilot at $10,000 to $40,000. Alice Labs lists pilot implementation at $80K to $300K. Both are the firms' own indicative ranges.8
Good
Time and materials
Alice Labs (2026) lists $80 to $900 an hour by firm type: nearshore $80 to $250, boutiques $250 to $600, Big Four $400 to $900. It puts senior hourly rates at $250 to $450.10
Retainer
Layer3Labs says $2,000 to $8,000 a month for small businesses and suggests budgeting 15 to 25 percent of the build cost per year for maintenance. Alice Labs lists managed AI operations at $5K to $25K a month.8
Outcome-based
Usually a base fee plus a share of saved cost or new revenue. Slalom lists Agentic Managed Services under outcome-based contracts. One blog says such deals are under 8 percent of contracts, citing SPI Research (unverified).11
Kodec, as one option
Kodec is a technology and operations firm for owner-run companies in home services, real estate, mortgage and law, in the $1M to $20M range. It builds growth systems, internal tools, AI agents and automation, and private AI systems that run on a company's own data. Kodec builds a working wireframe first, so you see the job before you commit. Every dollar and hour is tracked against results, which is the same table shown above. Kodec is not the right call for a company that needs a multi-year program across many sites. Accenture, Deloitte or BCG X fit that better.
What we found
Verified terms
What each badge requires, in plain words
Anthropic is the only program here that publishes its bar. Select needs 10 active certified people, 2 customers in production in the last 12 months and 1 public story. Preferred needs 100, 15 and 3. Global Premier needs 1,000, 100 across three regions, 15 public stories and a joint business plan. Anthropic says it verifies standing every quarter, and that a smaller firm climbs by growing its certified bench. Its directory lists 17 Global Premier, 42 Preferred and 295 Select partners.
So a Select badge means roughly this: ten people passed exams and two customers run Claude in production. That is a real floor. It is also a low one. Certifications belong to people, so ask whether those people will work on your system. Anthropic itself says a tier measures built practice, not outcomes.
OpenAI's network is newer. It has Select, Advanced and Elite levels and three specializations, Codex, Cybersecurity and Agents. The one source we found says OpenAI has not published numbers, and that its goal of 300,000 certified consultants is a target, not a count. Check openai.com/business/partners before you rely on a badge.
AWS validates a firm's expertise for its Generative AI Competency. Third-party summaries say Microsoft scores points for deal performance, certifications and customer success, and Google ranks by customer outcomes. We did not find thresholds for Microsoft or Google. Ask any firm to show you the page on the vendor's own site, and ask for two customers you can call.13467
Original math
How to score the 12 questions
Score each answer 0, 1 or 2. 0 means no answer or a vague one. 1 means a general answer. 2 means a specific answer you can check. Total out of 24. Under 14, keep looking. 14 to 19, ask follow-ups. 20 and up, you have a serious candidate. The red flags listed with the questions come up across the guides we read. All of those guides are written by firms that sell implementation, and they overlap on the same few items.131214
Playbook
What implementation includes, and who writes the guides
Seven parts: discovery (which job, what it costs now), data (what the system reads and how clean it is), build, integration with your tools, evals (tests on real examples from your business), rollout to your team, and run (monitoring, fixes, model changes). Ask each partner which parts are in the price. A quote that covers only the build leaves integration, evals, rollout and run for later bills.
We read the guides listed in Sources. Each one is published by a firm that sells AI implementation, including Leanware, Your Tech Club and Next Pillar. Treat any checklist, including ours, as the seller's. That is why this page tells you to measure results in your own numbers.141213
Questions to ask
- 01
What is the finished job, in hours saved and dollars?
A good answer: They fill in the table above with your numbers.
Protects you from: A demo that never turns into savings.
- 02
Name three systems you built that run in production today.
A good answer: Names, dates, and a customer you can call.
Protects you from: Unnamed clients and unchecked claims.
- 03
Who will actually do the work, and can I meet them before I sign?
A good answer: Names and hours per week.
Protects you from: A senior pitch and a junior build.
- 04
What will you look at in my data before you promise anything?
A good answer: They ask for a sample and say what is missing.
Protects you from: A promise made without reading your records.
- 05
Whose information does this use, where does it live, and which AI providers see it?
A good answer: A one-page answer.
Protects you from: Your customer data going to tools nobody told you about.
- 06
What do we own at the end?
A good answer: Code, prompts, settings and accounts are yours, in writing.
Protects you from: Being locked in.
- 07
How will it connect to the tools we already use?
A good answer: They name your systems and what each connection does.
Protects you from: A smart tool nobody opens.
- 08
How will you test it before real customers touch it?
A good answer: A set of real examples from your business, a pass mark, and a human check on risky actions.
Protects you from: Errors found by your customers.
- 09
How do we roll it out to the team?
A good answer: One team first, a trial period, a named owner on your side.
Protects you from: A tool your staff quietly avoid.
- 10
Who runs it after launch, and what does that cost per month?
A good answer: A named person and a price.
Protects you from: A system that decays by month nine.
- 11
What happens if we stop after the first phase?
A good answer: You keep everything built so far.
Protects you from: Pricing that punishes leaving.
- 12
When would you tell me not to build this?
A good answer: A real example. One agency's guide lists 'every recommendation leads to a build project' as a warning sign.
Protects you from: A partner who sells whatever you ask for.
Red flags
- A guaranteed result before anyone has looked at your business.
- A solution proposed before any process was examined.
- No one can say what the current work costs you today.
- Unnamed clients, or numbers with no source.
- No plan for who maintains the system after launch.
- Monitoring or support sold as an extra you only learn about later.
- Unclear ownership of code, prompts and settings.
- A promise of return in 90 days from someone who has not seen your data.
- Phase-gated pricing that makes leaving early almost as costly as finishing.
- A badge shown as the proof, with nothing else.
How we checked
We searched the five phrases above on 2026-11-14 and read the pages from firms that rank for them. We took program rules from Anthropic's own pages where they exist. OpenAI, Microsoft and Google details come from third-party summaries and are marked as such. Price ranges come from firms' own blogs and rate guides, not from quotes. We did not find published prices for Accenture, BCG X or Thoughtworks. We re-check this page monthly.
Questions
It takes you from a chosen job to a running system. That covers discovery, data, build, integration with your tools, testing, rollout to your team and ongoing run. Ask which of the seven parts are in the quote, because many quotes cover only the build.
No. A badge shows a firm joined a program and met its bar. Anthropic says its tiers measure built practice, not outcomes. Use a badge to build a shortlist, then ask for two customers you can call and a number for the finished job.
Firms rarely publish prices. Layer3Labs lists a single-workflow pilot at $10K to $40K, and Alice Labs lists pilot implementation at $80K to $300K. Alice Labs reports hourly rates from $80 to $900 depending on the firm. Treat all of these as third-party ranges and ask for a fixed price on the first step.
Count hours saved and dollars kept against the number written before you signed. Measure again 90 days after launch. If the partner will not write the number down, that is your answer.
Sometimes. If you have many sites, strict audit rules or a program across several departments, a big firm can be the better choice. If you need one process fixed with a price and a date, a smaller firm often fits better. Ask both for the same proof.
Sources
- 1.Anthropic: Claude Partner Network Services Track and Partner Hub accessed Nov 14, 2026
- 2.Anthropic: Claude Partner Network announcement accessed Nov 14, 2026
- 3.Claude service partners directory accessed Nov 14, 2026
- 4.Digital Applied: OpenAI Partner Network analysis (secondary) accessed Nov 14, 2026
- 5.OpenAI partner page (not read) accessed Nov 14, 2026
- 6.AWS: Generative AI Competency partners accessed Nov 14, 2026
- 7.flashdba: AWS, Microsoft and Google partner programs compared (secondary) accessed Nov 14, 2026
- 8.Layer3Labs: AI implementation partner pricing accessed Nov 14, 2026
- 9.Prometheus Agency: AI agency pricing and engagement models (2026-04-28) accessed Nov 14, 2026
- 10.Alice Labs: AI consulting rates 2026 accessed Nov 14, 2026
- 11.Sombra: best AI consulting firms accessed Nov 14, 2026
- 12.Your Tech Club: AI implementation partner guide (2026-10-30) accessed Nov 14, 2026
- 13.Next Pillar: choosing an AI partner accessed Nov 14, 2026
- 14.Leanware: how to choose an AI implementation partner accessed Nov 14, 2026
Where to go next
Want to see which of your own calls turned into paying customers? See how Leverage works.
Related answers
- Home servicesAI & operations
Is AI for HVAC companies worth it in 2026, and where does it pay off?
Updated Nov 14, 2026
- All industriesAI & operations
What can AI agents for business operations actually do?
Updated Nov 14, 2026
- LawAI & operations
What can AI for law firm management do for intake, billing and operations?
Updated Nov 14, 2026