For a business under about 25 people that wants ChatGPT working on a few everyday jobs, hire an independent consultant on a fixed scope, or try it yourself first. A firm only earns its higher fees when you need security sign-off, links to other systems, or several teams switched on at once.
That 25-person line is my rule of thumb, not a published figure. What matters more is how many systems ChatGPT must touch and how much written assurance your clients or insurer expect. Note too that "OpenAI consulting firm" is not a protected title. OpenAI does run the OpenAI Partner Network, with a public partner locator that listed 73 organisations when I checked in November 2026, and the client stories on its partner page are large enterprises working with firms such as Accenture, BCG and Bain. So a partner badge tells you a firm is enrolled, not that it suits a ten-person office; if a firm claims it, find the firm in the locator. And any proposal built around custom GPTs is already out of date: OpenAI is retiring them, and existing ones are scheduled to stop running on 11 December 2026.
A ten-question test: do it yourself, hire solo, or hire a firm
Score each question 0, 1 or 2 and add the points. A higher total means more reason to pay for help. The questions measure two things: how clear you already are about the jobs, and how much risk or plumbing sits around them.
| Question | 0 points | 1 point | 2 points |
|---|---|---|---|
| 1. Is there a named owner with two hours a week for a month? | Yes | Maybe | No |
| 2. Has anyone used ChatGPT for work for three months or more? | Two or more people, regularly | One person, casually | Nobody |
| 3. Can you list the first three tasks, with a real sample of each? | Yes | Vague ideas | No |
| 4. Must ChatGPT connect to email, a CRM, accounting or file storage? | No | One system | Two or more |
| 5. Will client or staff personal data be typed in? | No | Sometimes, and it can be redacted | Routinely |
| 6. Do clients, a regulator or your insurer expect written AI controls? | No | Possibly | Yes |
| 7. How many people will use it in the first 90 days? | 1 to 5 | 6 to 25 | 26 or more |
| 8. Is a deadline driving this within eight weeks? | No | A soft one | A hard one |
| 9. Have you already tried it yourself? | Not yet | Yes, results are patchy | Yes, and it stalled |
| 10. Would a failed setup cost money or client trust, not just a wasted week? | No | Some | Yes |
Read the total like this. From 0 to 7, do it yourself for four weeks and buy two or three hours of advice only if you stall. From 8 to 13, hire an independent consultant on a fixed scope. From 14 to 20, ask firms for quotes, but also collect one independent quote so you can see what the extra money buys.
Take a twelve-person HR consultancy that scores itself. The owner question is a maybe (1). One partner uses ChatGPT casually (1). The first three tasks are still vague (1). It would need email and a shared drive connected (2). Employee case notes are typed in routinely (2). Two clients have asked about written controls (1). Twelve people will use it (1). A client review in seven weeks is a soft deadline (1). A pilot gave patchy results (1). A failure would cost some client trust (1). The total is 12, which points to an independent consultant on a fixed scope.
The useful part is what moves the score. If that consultancy's largest client demanded a written AI policy, question 6 becomes a 2. If a second office doubled the user count, question 7 becomes a 2. Those two changes alone push the total to 14, and the answer flips towards a firm. The test is driven by plumbing, assurance and scale, not by how ambitious the jobs are.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowFees: what each route costs, using published bands
Doing it yourself costs seats and your own time. As of November 2026, ChatGPT Business Standard seats list at $20 per user per month on annual billing or $25 monthly, for teams of 2 to 200 users. Check OpenAI's pricing page before you budget, because plan names and prices change.
For independents, the closest published figures come from Contra's page on hiring Zapier freelancers, who are automation specialists: $30 to $150 an hour, $500 to $1,500 for simple automations, $2,000 to $10,000 or more for complex ones, and retainers of $1,000 to $5,000 a month. Those are estimates for automation work rather than ChatGPT coaching, so treat them as a floor and ceiling, not a quote.
For firms, Clutch's 2026 AI pricing guide reports most AI development companies charging $24 to $49 an hour, reviewed projects ranging from $10,000 to $49,999, and an average project cost of about $120,000 over roughly ten months. Clutch splits its hourly bands by where a firm is based, so the hourly figure alone tells you little about a firm with senior staff in your time zone. The project floor tells you more, and the bottom of Clutch's $10,000 to $49,999 range is a fair working floor for a firm engagement: that is what you are agreeing to before anyone opens your workspace.
| Route | Basis for the sum | Year-one cost for the twelve-person HR consultancy |
|---|---|---|
| Do it yourself | 12 seats at $20 a month on annual billing, plus about 16 hours of the owner's time | $2,880 in seats (12 × $20 × 12); $3,600 if billed monthly at $25 |
| Independent consultant | 20 hours at $100 an hour, inside Contra's $30 to $150 range | $2,880 + $2,000 = $4,880 |
| Consulting firm | Project floor of $10,000, the bottom of Clutch's reviewed-project band | $2,880 + $10,000 = $12,880 or more |
These are illustrative sums, not quotes. The point is the gap: for the same three jobs, the firm route costs more than two and a half times the independent route in year one, and the seats cost the same on every route. If the extra $8,000 buys security paperwork or integrations you really need, it can be money well spent. If it buys a slide deck and a project manager, it cannot.
Who does the work, and what each contract usually says
The biggest practical difference is the person who configures your workspace. In a firm, the person who pitches may not be the person who delivers, and some of the delivery may be subcontracted. In a solo practice, the person you meet is the person who does the work, which is a strength, but there is no second pair of eyes and no cover if they fall ill mid-project.
| Point | Typical firm | Typical independent |
|---|---|---|
| Fee shape | Project floor, discovery phase, then build | Fixed scope or hourly, sometimes a small retainer |
| Who delivers | A team; the pitch lead may not be in the room | One person |
| Contract | Firm's master agreement, with its own liability cap and terms | Shorter agreement, often negotiable, or your template |
| Default on ownership | Check carefully: some firms keep rights in reusable assets | Usually agreed case by case, so you must ask |
| Security paperwork | Often has questionnaires and policies ready | Varies; ask what they can produce |
| Continuity | Colleagues can step in | Ask who covers illness and whether they hold professional indemnity insurance |
| Best fit | Several teams, integrations, formal assurance | Three to five jobs, one workspace, a team under about 25 |
These are tendencies, not rules, and they overlap: some firms are three people and some freelancers work through a network. The sibling tutorial on consultant, agency or freelancer covers the generic comparison. What matters here is the question to put to either: name the person who will configure my workspace, and tell me how many hours each named person will spend.
Putting prompt and workflow ownership in writing
In a ChatGPT setup, "prompts and workflows" means concrete things: the instructions you attach to each Project, the reference files, the library of tested prompts, any template documents, the written process for each job and the training material. Each of those should be named in the contract as a deliverable that belongs to you once you have paid for it.
OpenAI's terms hand you whatever rights it has in the outputs, "if any", so the contract with the consultant is where your real protection sits. Four points matter most.
- Built in your workspace. The work happens under your administrator, not inside the consultant's personal or trial account.
- Assigned on payment. Everything made for you transfers to you, while the consultant keeps generic know-how and tools they brought with them.
- Master copies outside ChatGPT. ChatGPT Business workspaces have no self-service data export, and chats of removed members are kept indefinitely with the owner unable to see them. So the final instructions and prompts must also sit in a document your firm controls.
- No shared logins. OpenAI's terms ban sharing account credentials. The consultant gets their own invited seat, you agree who pays for it, and you remove it at handover. When you do, OpenAI reassigns any Projects that seat created to a workspace owner, but the consultant's own chats and files are not transferred, so anything worth keeping must already sit in a Project or in your document.
An illustrative clause you could adapt, to be reviewed by a lawyer before use:
Deliverables. The Consultant will build all Projects, project instructions,
reference-file sets, prompt libraries, templates and process documents
inside the Client's own ChatGPT workspace, and will also supply a copy of
each in an editable document. On payment of the fees due, all rights the
Consultant holds in those deliverables pass to the Client. The Consultant
keeps its general methods and any tools it owned before the engagement.
The Consultant will not use shared credentials, will list every account it
has been given, and will confirm in writing at handover that its access
has been removed.The sibling tutorials on who owns the workflows a consultant builds and nine contract clauses to check go deeper on the legal side. Neither replaces a lawyer reading your actual contract.
What went wrong when the build sat in the consultant's account
A small recruitment agency paid for four days of setup. The consultant built the Projects in their own workspace "to move faster" and sent screenshots at the end. When the agency tried to copy the instructions into its own account, half of the reference files had been renamed and two prompts depended on settings it could not see. Rebuilding took the agency's office manager about 14 hours, and a month later nobody could say which version was correct.
The fix costs nothing: put "built in the Client's workspace, administered by the Client" in the scope, and check on day one that the consultant's seat exists in your workspace. This is an illustration, not a named client, but the failure is easy to reproduce whenever the work lives in someone else's account.
A scoped ChatGPT setup checklist to hand to either
This is the scope I would send to a solo consultant and a firm alike, so that quotes are comparable. It assumes three everyday jobs for a team under 25 and a Business workspace. The hours are my estimates for the consultant's side only.
| Step | Deliverable | Done when | Hours |
|---|---|---|---|
| 1. Plan and seats | Decision note on plan, seat list, who pays for the consultant's seat | Every user has their own login and the two-seat minimum is met | 1 |
| 2. Admin and data settings | Named administrator; training default confirmed; apps (formerly connectors) switched on only where a job needs them | The administrator can show the settings screen without help | 2 |
| 3. Three Projects, one per job | Project instructions plus 3 to 8 reference files each (Business allows up to 40 files per Project) | A colleague gets usable output first time on five sample inputs | 6 |
| 4. Prompt library | 8 to 12 tested prompts, each with a sample input, a sample output and a "check this" note, saved in your own document store | Library opens without a ChatGPT login | 3 |
| 5. Use rules | One page on what never goes in and who checks which outputs | Every user has read it and confirmed in writing | 2 |
| 6. Staff session | A 60 to 90 minute live session, recorded, plus a 30-day check-in | Each person completes one real task unaided | 3 |
| 7. Measure | Time taken before and after on ten samples of each job | A sheet of numbers exists | 2 |
| 8. Handover | Seat list, administrator, renewal dates, documents, consultant seat removed | Checklist signed by both sides | 1 |
That adds up to 20 hours, or $2,000 at $100 an hour. Put what is out of scope in the same document: building custom GPTs (being retired), building agents, connecting ChatGPT to systems through the API, legal review of your policies, and anything that needs a second workspace. A firm will often want to price steps 1 and 2 as a separate discovery phase; that is reasonable only if the phase has its own written deliverable.
Step 7 deserves a defence, because it is easy to cut. Without before and after times you cannot tell a good setup from a pleasant afternoon. Time ten real cases of each job by hand, then time the same ten through the new Project, and record how many outputs needed a rewrite. The handover checklist sibling lists what else to collect before the consultant leaves.
Proposal red flags that show the advice is out of date
ChatGPT changes monthly, and a proposal reveals whether the author keeps up. These are the flags I would look for in any quote, solo or firm.
- Custom GPTs as a deliverable. OpenAI is retiring them, and existing ones are scheduled to stop running on 11 December 2026. The replacements are Projects for shared context, and OpenAI's migration for GPTs you already have, which turns a GPT's instructions into a skill inside a plugin.
- "Agent mode" as the headline feature. OpenAI now sells longer multi-step tasks under ChatGPT Work, which its Business pricing page lists as a plan feature, so a quote should name the feature it means and show it running in your workspace.
- "We will train ChatGPT on your documents." A Project supplies your instructions and reference files for ChatGPT to work from; nothing in that setup retrains the model. Business plans also do not train on your business content by default, so the claim is either loose wording or a misunderstanding.
- Shared logins or a request for the owner's password.
- A partner or certification badge with no document behind it. Ask to see the firm's entry in OpenAI's partner locator, or the certificate number.
- A lump sum with no acceptance test. "Done" should mean a named person completes a named task unaided.
- Savings promised in advance. No one can know your time savings before measuring ten real cases.
The tutorial on red flags to walk away from adds a longer list that applies beyond ChatGPT.
Questions to put to a firm and to a solo consultant
Ask both kinds of supplier the same questions, so the answers compare. What follows each question is what a good answer sounds like.
| Question | A good answer sounds like |
|---|---|
| Who will configure my workspace, and how many hours will each named person spend? | Names, roles and hours, with any subcontracting disclosed |
| Can I see a sample handover pack? | A redacted pack with instructions, prompt library and a seat list |
| What do you need from me each week? | A realistic figure, such as two hours from one owner and sample inputs by day three |
| What is not in scope? | A written list, including custom GPTs and API work |
| What happens if a ChatGPT feature changes mid-project? | An example of a past change and how they handled it |
| Do you hold professional indemnity cover, and for how much? | A certificate on request, with the limit stated |
| Do you earn anything from the tools you recommend? | A plain no, or a plain disclosure |
| How will we prove it worked? | Before and after timings on real samples, agreed up front |
Pay attention to how each supplier answers the fourth question. A supplier who lists exclusions without prompting has done this before. One who says "we'll figure it out as we go" is selling you their learning time. If you want a longer list, twenty questions to ask an AI consultant first covers most cases.
A paid half-day as the tiebreaker
When two suppliers look equal on paper, pay each finalist for a half-day on one real job. At $100 to $150 an hour, four hours costs $400 to $600, a small amount against a project worth thousands. Give each the same sample inputs and watch how they work in a shared Project.
Score them on five things. Did they ask what data the job touches before typing anything? Did they show you a failure case and how to catch it, or only polished successes? Did the output improve after a second round of instructions? Did they leave behind a document you can use without them? Did the hours they billed match the hours they said? A half-day also reveals the gap that matters most between firms and individuals: whether the person in the room is the person you would get.
I should name my own interest: I am an independent consultant, so weigh this advice accordingly. If you want a second opinion on the scope before you send it out for quotes, that is what my AI implementation consultation covers.
Questions small firms ask before paying for ChatGPT help
Is there an official OpenAI-approved consultant list for small firms?
OpenAI runs the OpenAI Partner Network and publishes a partner locator on its business site. The client stories on its partner page are large enterprises, so few listed firms aim at small offices. If a firm claims partner status, look it up in the locator before you pay a premium for the badge, and judge it on the same scope and fees as anyone else.
Can a consultant use my ChatGPT login to build things?
No. OpenAI's terms ban sharing account credentials, so each person needs their own login. Invite the consultant as a member with only the access the job needs, agree who pays for that seat, and remove it at handover. Keep the master copy of every instruction and prompt in a document you own.
Should I buy custom GPT work if the quote is cheap?
No. OpenAI is retiring custom GPTs: existing ones are scheduled to stop running on 11 December 2026, and OpenAI's migration route turns each one into a plugin. Ask for ChatGPT Projects with written instructions and reference files, plus a prompt library stored outside ChatGPT, which survive product changes better.
How long should a small ChatGPT setup take?
The scope in the checklist above adds up to about 20 hours of consultant time, spread over two or three weeks, plus an hour or two a week from your internal owner. That figure is my estimate for three everyday jobs, not a published benchmark. A quote measured in months for the same scope deserves questions.
Further reads
- What an AI Consultant Needs From You: Access, Data, and Time — The access, data and time a consultant will ask you for.
- What a Typical AI Consulting Engagement Looks Like, Week by Week — What a typical engagement looks like, week by week.
- How Much Does an AI Consultant Cost for a Small Business? — What small businesses typically pay at each level.
- Should You Sign an NDA Before Sharing Data With an AI Consultant? — Whether and when to sign an NDA first.
- How to Negotiate an AI Consulting Quote Without Cutting Corners — How to push back on a quote without cutting corners.
- How to Vet an AI Consultant's Case Studies and References — Checking the work a consultant says they did.
- Hire an AI Consultant for an Accounting Firm: Scope, Cost, Red Flags — The same hiring decision for an accounting practice.
- Hiring an AI Consultant for an Insurance Agency: Scope and E&O — Scope and insurance questions for a regulated agency.
- AI Consultant for Nonprofits: What to Hire For and What to Pay — Day rate, fixed fee or retainer: what a nonprofit AI consultant delivers, what it costs, and how to write a brief that stops scope creep.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI pricing and help pages for ChatGPT Business, Projects and custom GPT retirement; Clutch AI pricing guide; Contra hiring page for automation freelancers; OpenAI Partner Network page and partner locator; OpenAI terms of use; OpenAI help articles on data export and removed workspace members.