Six Red Flags in an Automation Proposal That Should Stop the Meeting

You are reading a set of claims, and the only ones that matter are the ones that can be checked later. Everything else is decoration, however good it looks.

That sounds obvious until you try it on a real document. Take any automation proposal, go through it line by line, and ask of each sentence: on what date could somebody prove this wrong? Most sentences will not survive the question. "Streamline your client onboarding" cannot be false. "Reduce manual handoffs" cannot be false. "Leverage AI across your delivery workflow" cannot even be measured, which is a different problem wearing the same coat.

The six flags below are all versions of one failure. Each is a place where a number, a date, a name or a price should be, and something unfalsifiable is sitting there instead. None of them proves the person across the table is dishonest, and several of them are habits picked up from a whole category that writes this way. But each one is a question the document should already have answered, and asking it costs you a single email.

"Save 20 hours a week." "Cut turnaround by 60%." "Three times more output from the same team." Ask which workflow, whose hours, and measured against what baseline, and watch what comes back.

A scoped version of the same claim sounds completely different. It names one process, it names who does it today, it says how long that takes now and how that was established, and it says what would count as the number being hit. That version can be wrong on a specific Tuesday, which is exactly why nobody writes it unless they intend to be held to it.

The unscoped version is not rare and it is not always deliberate. It has become the house style of the whole category. Gartner's June 2025 forecast put more than 40% of agentic AI projects on track to be canceled by the end of 2027, and named the mechanism in the same release: agent washing, which is rebranding an existing assistant, chatbot or RPA script as something agentic without changing what it does. Of the thousands of vendors claiming agentic capability, Gartner reckoned only around 130 actually had it.

There is also a floor under how far a number can drift before it stops being marketing. In April 2025 the FTC ordered Workado to back up its accuracy claims for an AI content detector advertised as 98% accurate. On general-purpose content the Commission said the real figure was about 53%, which is close enough to a coin flip to be worth nothing, and the final order landed that August. The lesson for a buyer is not that everybody is Workado. It is that a percentage in a pitch deck carries no evidentiary weight at all until somebody says how it was measured, and a regulator had to say so out loud.

If a seller cannot price a defined scope, the scope is not defined. That is the whole flag, and it is the one I would put first if the list were ordered by how much money it saves you.

The standard defense is that every project is different. Some of that is true. All of it is also true of every plumber, every accountant and every managed service provider you have ever hired, and they all manage to publish a rate. What the defense usually protects is the ability to price you rather than the work, which is a legitimate commercial strategy and a terrible thing to be on the receiving end of.

Buyers have been saying this out loud for years. TrustRadius surveyed 1,604 technology buyers in 2023 and found that software pricing not being on the website was the single biggest reason they would be less likely to buy, at 54%, while 72% said transparent pricing made them more likely to. Transparent pricing has topped their buyer wish list every year they have asked. The market has been telling sellers what it wants for four years and most of the category still routes to a discovery call.

So here are ours, because a post about missing prices that does not print its own is doing the same trick. A one-off starter build, one scoped automation shipped and handed over, is $1,500 to $2,500 fixed. A two-week production sprint that puts a full workflow live is $5,000 fixed. Retainers start at $5,000 a month, and that is a floor for ongoing build-and-run work rather than a minimum spend to talk to us. Model APIs, tools and subscriptions are billed to your accounts and never marked up. The full ladder is on our pricing page, and it has not moved all year.

Ask who they have turned down in the last six months, and why. A shop that has never said no to anybody has no opinion about what it is good at, and you are about to buy the opinion.

This is the flag the whole category skips, and it is skipped for an obvious reason. An anti-fit statement costs the seller leads. It is also the cheapest possible proof that somebody has judgment, because refusing revenue is the one signal that cannot be faked in a deck. Watch for the difference between a real answer and a polite one. A real answer names a situation: too early, process not written down, one connector would have fixed it, the internal person who wanted it was leaving. A polite answer says they focus on clients who are ready to commit, which is not a criterion, it is a compliment about the people who said yes.

Ours, since asking a question you have not answered yourself is cheap. Under about five people, do not buy a build, because the same task is not repeating often enough yet to pay one back. If the process has never been written down, write it down first, because nobody can automate the version of a workflow that three people run three different ways. If one connector between two tools fixes the whole thing, buy the connector, and anyone charging thousands to install it is charging you for the install. And if you have an ops person who enjoys this stuff and has slack in their week, give them the time instead, because they will build something worse and maintain it forever, which regularly beats something better that nobody in the building understands. We laid the same argument out at length in the guide to what an agency automation consultant actually does, including the three endings that involve hiring nobody at all.

Every pilot needs one line that is either true or false on the last day, written before the first day. If the proposal does not contain that line, write it yourself and send it back.

It looks like this: "On day fourteen, the intake form submits, the brief is drafted, and the account manager approves or rejects it inside the tool, for 90% of the jobs that came in that week, without anyone touching a spreadsheet." That is a sentence with a date, a threshold, a population and a failure condition. Compare it to "deliver a working proof of concept demonstrating the value of automation in your intake process", which cannot fail, and therefore cannot succeed either.

The cost of skipping this is measurable. S&P Global Market Intelligence surveyed more than 1,000 organizations across North America and Europe and found that the share scrapping most of their AI initiatives had jumped to 42% in 2025 from 17% the year before, and that the average organization abandoned 46% of its AI proofs of concept before any of them reached production. Nearly half the pilots die in the gap between something impressive happening in a demo and something boring happening every Tuesday. A pilot with an acceptance line closes that gap on paper before you have paid for it.

One more thing to insist on: the line has to be checkable by you, not by them. "Model accuracy above the agreed threshold" is a sentence only the builder can evaluate. "The account manager did not open the spreadsheet once this week" is a sentence you can check by asking the account manager.

Ask what happens to the price and the date when the scope moves, because it will move. A proposal that has no answer is not a fixed-price proposal, it is a fixed-price proposal until the first awkward conversation.

Scope movement is the normal condition of this work rather than a failure of planning. PMI's 2018 Pulse of the Profession found 52% of projects completed in the preceding year had experienced scope creep, up from 43% five years earlier, and that is across every kind of project run by organizations with actual project managers in them. An automation build touches systems nobody has documented, run by people who each do one step slightly differently. Something will be discovered in week one.

What you want written down is not a promise that it will not happen. It is the mechanism: who can request a change, who prices it, whether the original date moves or the original scope shrinks, and what the smallest unit of extra work is. A shop that has done this before has that paragraph already written, because they got burned once and added it. A shop that has not will tell you they are flexible, which means the negotiation happens later, when you have already paid a deposit and have less leverage than you have right now.

Look for the line that prices month two. In most proposals there is not one, and the silence is not an oversight, it is the shape of the product.

A build has a completion date. A running system has a bill: model and API usage, the platform seats, the person who notices when a step fails at 3am, the hour a month that goes into a vendor changing an endpoint. Ask who gets paged, on what channel, inside what response time, and what that costs. Then ask whose accounts the API usage is billed to, because that answer tells you whether you are buying a system or renting access to one.

The budgeting failure here is well documented. Gartner's guidance on optimizing generative and agentic AI costs puts at least 50% of these projects over their budgeted cost through 2028, and it blames poor architectural choices and a lack of operational know-how rather than the price of tokens. Architecture and operations get decided during the build. They get billed forever.

The end of that road shows up in the returns. Gartner has since found 74% of organizations breaking even or losing money on their AI investments, against an average of $1.9 million just to get a deployment started, before anybody is trained or anything is managed. Those are not stories about bad luck. They are stories about the second year of ownership arriving unpriced.

Our own build log has the specific version of this, with the failure points and the retry counts from systems we run: what actually breaks in production automations and roughly how often. The short version is that a system with nobody watching it is not a system, it is a scheduled surprise.

A checklist that only adds reasons to worry makes you slower without making you righter. Three things get treated as warnings and should not be.

No case study in your industry. Agency operations are agency operations, and intake, briefs, approvals and reporting look far more like each other across industries than the vertical landing pages suggest. What matters is whether they have automated a workflow shaped like yours, not whether the logo above it sold the same thing you sell. Ask about the shape of the process instead and you will get a much more useful answer.

The price moving after they see your systems. This is fine as long as the range came first and the change arrives with a reason attached. "It is $5,000 for a workflow of this shape, and yours has a legacy CRM with no API, which adds a week" is a seller doing their job. The flag is not the revision. The flag is a first number that never existed, so there is nothing to revise from.

Them saying no to something you asked for. People read this as inflexibility and it is usually the opposite. Somebody who tells you the thing you asked for is a bad idea, and says why, is showing you what the next twelve months of disagreement will feel like. It is a much better signal than enthusiasm, which is free. The version to worry about is the shop that agrees to everything in the room and then defers all of it to a change order later.

Do it once, in order, with a pen. Five of the six flags are lines that should be sitting in the document, so you can check them without talking to anybody. The anti-fit one is the only question you have to ask out loud.

Read for the price first, because if there is no number and none arrives when you ask, the rest of the reading is unpaid work. Then read for a named workflow, then for the acceptance line, then for the change-order paragraph, then for the month-two cost. Every one that is missing becomes one email with one question in it. Send them all at once rather than in a sequence, because a sequence gives everyone a week and tells you nothing.

Then measure the reply, not just the content of it. If the answers come back inside two working days with numbers in them, you have learned that this is roughly the response time you will get when something breaks in production. If they come back in nine days, or come back as a call invitation instead of an answer, you have also learned something, and it cost you nothing to find out. When all five lines are there, buy the smallest version of the work that exists rather than the whole thing. If you want the artifact-level version of this, we wrote a separate walkthrough of what to ask a builder for when you cannot read code, which is the same instinct pointed at the person rather than the document.

Everything above is a standard we would rather be held to than write about, so here is the version of us that you can check.

The prices are published and have not changed all year. A starter build is $1,500 to $2,500 fixed, one scoped automation shipped and handed over. A two-week production sprint is $5,000 fixed, one full workflow live, with handover documentation included. Retainers start at $5,000 a month for ongoing build-and-run work, cancel monthly, and you keep everything built either way. Third-party costs, the model APIs and the tool subscriptions, are yours directly and never marked up, which also means we have no reason to route your work through the expensive model.

The anti-fit list in flag three is not a rhetorical device, it is the list we actually run people through, and the free audit is where it gets applied. The audit maps where your hours currently go and ranks what is worth automating, and one of the available rankings is nothing, not this quarter. That map is yours to keep either way, which is the only reason it can honestly come back with that answer. An audit you have paid for tends to find work.

What we will not put in writing is a guaranteed outcome number before we have seen your systems, and that is the one place we will look worse than a competitor on a spreadsheet comparison. Somebody will always promise you a percentage. The gap between the two proposals is not capability, it is which of us is willing to write down something that cannot be checked. We would rather write the acceptance line and be held to that, and we build evaluation suites around agents for the same reason: a system that nobody measures is a system nobody can defend.