Short answer: you cannot compare AI automation quotes until you have made them comparable. Write the scope yourself, define one unit of work, then ask every supplier to price that same unit. Everything else in this piece is how to do that in an afternoon.
The buying question is covered elsewhere on this site: what business automation costs and what an AI consultant costs both answer the price question directly. This piece assumes the quotes are already on your desk and you cannot tell which one is the good deal.
Why three quotes for one job look nothing alike
Ask three suppliers to automate the same workflow and you get three documents with almost nothing in common. One is two pages and prices a pilot. One is fourteen pages and prices a platform. One prices people by the day and leaves the outcome open.
None of them is being dishonest. They are answering different questions, because nobody told them which question to answer. The brief said "we want to automate our inbound enquiries", and each supplier filled the gap with the version of that job they are best at selling.
This is why the usual advice, get three quotes and take the middle one, fails badly on AI work. The middle quote is the middle of three different jobs. You are not choosing a price. You are choosing, without meaning to, which of three scopes you are about to buy.
Step one: write the scope yourself, on one page
Before you read another proposal, write down what you actually want, in your own words, in about a page. Not a specification. A description a new employee could follow.
- The workflow as it runs today. Who touches it, in what order, using which systems. Include the ugly parts, the spreadsheet, the shared inbox, the person who remembers the exceptions.
- The trigger and the finish line. What starts the work and what state counts as done. "A call comes in" to "the meeting is in the calendar and the customer has a confirmation" is a finish line. "Better customer service" is not.
- The volume. How many times a month this happens today, and what it looked like in your busiest month last year.
- What must not change. The system of record you are keeping, the tone customers expect, the regulatory constraint you cannot negotiate away.
Send that same page to every supplier. It costs you an hour and it converts three incomparable documents into three answers to one question. It also does something quieter: a supplier who reads your page and comes back with a better description of your own process than you wrote has told you something useful about themselves.
Step two: pick one unit of work and make everyone price it
A monthly figure means nothing on its own, because AI systems bill by use. The comparable number is the cost of one unit of work: one answered call, one processed invoice, one qualified lead, one resolved ticket.
The vendors underneath price this way too, and they publish it. OpenAI and Anthropic both meter text in and text out, Twilio meters connected call minutes and n8n meters workflow executions. A supplier quoting one flat monthly figure is either absorbing that variability or has not looked at it.
Ask each supplier for two numbers against your stated volume: what this costs at today's volume, and what it costs at three times today's volume. The gap between those two numbers is the shape of the pricing, and it exposes more than any headline figure. A quote that barely moves between the two is mostly fixed cost. One that triples is mostly usage. Neither is wrong, but they suit very different businesses.
The running side deserves its own attention, because build quotes routinely omit it. We took that apart line by line in what an AI system costs to run. For the comparison sheet, all you need is the supplier's own answer, in writing, to a question they cannot dodge: what will we be paying you and your vendors in month six.
Step three: the eight lines that decide the real total
Read every proposal for these eight things and mark each one present, vague or absent. The pattern of blanks tells you more than the prices do.
1. Scope, in your words
Does the document describe your workflow, or does it describe the supplier's product with your logo on it. A proposal that could be sent to another company by changing the name has not been written for you.
2. Integrations and who provides access
Every system the automation touches should be named, with a note on who is responsible for access and what happens if that access is refused by your own IT. Unnamed integrations are where timelines die.
3. What it costs to run, separately from what it costs to build
Two figures, never one. If the proposal shows a build price and no monthly figure, the monthly figure is not zero. It is unquoted, which means it is yours to discover.
4. Whose name the vendor accounts sit in
Model keys, telephony, the automation platform, hosting. Accounts in your name mean you can see the real spend and change supplier without a rebuild. Accounts in the supplier's name mean neither. This line costs nothing to get right at the start and is expensive to fix later.
5. A written acceptance test
How will both parties know it worked. Not a feeling, a test: this many real cases pass end to end, exceptions route to a named person, the log shows what happened. Without it, "done" becomes a negotiation between two people who each remember the meeting differently.
6. A timeline with milestones, not a duration
"Six weeks" is a wish. "Access confirmed by week one, first workflow live in staging by week three, ten real cases passed by week five" is a plan you can hold someone to, and a plan that fails loudly and early rather than quietly and late.
7. What happens when scope changes
It will. The question is whether the document says how. A named rate, a change process and someone empowered to approve it beats silence, because silence resolves into an invoice.
8. Handover and exit
What you own at the end, where the documentation lives, and what it takes for another supplier or your own team to run it. A system you cannot inherit is a subscription with extra steps.
Five phrases that should slow you down
None of these means the supplier is bad. Each one means ask the next question before you sign.
- "Intelligent" or "smart". Ask which part actually calls a language model and which part is ordinary software moving data between systems. Often the useful half is the plumbing, which is fine, but you should know what you are buying.
- "Pilot". Ask what turns the pilot into production, who pays for that, and what the acceptance test is. Pilots that were never scoped to graduate are the most common way this money disappears. We wrote up the pattern in why AI pilots fail to reach production.
- "Custom model". Ask why an off the shelf model plus your own data will not do the job. Sometimes the answer is good. When it is not, you have just bought a maintenance obligation.
- "Per seat". Automation replaces work, not seats. Seat pricing on a system that runs without a human in the loop means the price is untethered from what the thing does.
- "Fully autonomous". Ask who reads the exceptions. Every working system has someone doing that. A supplier who says nobody does has either not run one or is not counting your staff time.
The comparison sheet, in ten minutes
Open a spreadsheet. One column per supplier. One row for each of the eight lines above, plus build price, month one running cost, running cost at triple volume, and cost per unit of work. Fill it in from the documents only, not from what you remember someone saying on a call.
Two things fall out immediately. First, the blanks cluster: one supplier will have five empty cells and it will not be the expensive one. Second, the cheapest build price and the cheapest total almost never belong to the same column, because the cheap build is usually the one that stopped at the demo and left error handling, monitoring and the handover for later.
Then go back to each supplier with their own blanks. What you learn from how they fill them is worth more than the original document. Fast, specific, in writing is a good sign. A meeting invitation instead of an answer is also a sign.
When only one supplier answers properly
This happens more often than the tidy three way comparison, and it feels like the process failed. It did not. A supplier who can state the running cost, put the accounts in your name and write down the acceptance test has run a system in production before, and that is exactly the thing you were trying to find out.
At that point the honest question is no longer which quote is cheapest, but whether the job is worth doing at all at that price. That is a different exercise: the cost of leaving the workflow alone, which we cover in the cost of inaction. Compare the good quote against doing nothing, not against two documents that were never really competing.
The short version
Write your own one page scope. Define one unit of work and make everyone price it at today's volume and at triple. Score the eight lines. Treat blanks as information. Then read the prices, last, when they finally mean the same thing.
Frequently asked questions
How do I compare AI automation quotes fairly?
Normalise them before you read the prices. Write your own one page scope, define one unit of work such as one answered call or one processed invoice, and ask every supplier to price that same unit. Then compare scope, integrations, running cost, account ownership, acceptance criteria, timeline and change terms.
What should an AI automation proposal contain?
The workflow it replaces described in your words, the systems it touches and who provides access, what it costs to run as well as to build, whose name the vendor accounts sit in, a written acceptance test, a milestone timeline, and what happens when scope changes.
Why do AI quotes for the same job differ so much?
Because they are usually not the same job. One supplier prices a demo, another prices a production system with error handling, monitoring and a handover. The scope gap hides inside summaries that look alike, which is why you write the scope instead of letting three suppliers each invent their own.
Is the cheapest AI quote usually the most expensive one?
Often, because the cheap version stops at the demo. The work that makes a system survive contact with real customers is the part that gets dropped to hit a number, and you pay for it later as a rebuild or as staff time nobody counted.
What if only one supplier answers the questions properly?
That is a result, not a failure. Treat the completeness of the answer as evidence that the supplier has run one of these before, and compare that quote against doing nothing rather than against two documents that were never really competing.
If you would rather not run this exercise alone, that is roughly what the free audit does from the other side: we map the workflow, name the systems it touches and tell you what the work is worth before anyone quotes it. Our consultancy overview explains how we scope, the audit first model explains why we start there, and you can book the free audit and bring the quotes you already have.