Don't sell AI outcomes until you know the baseline

Outcome pricing sounds like the honest way to sell AI agents.

“Only pay us when it works.”

Clean. Confident. Very buyer friendly.

Also dangerous, if nobody knows what “works” means yet.

That is the bit most AI agencies will skip because it ruins the pitch. They want to jump straight to the sexy commercial model: percentage of revenue created, percentage of cost saved, pay per lead, pay per booking, pay per qualified account, pay per generated asset that ships.

Some of that will work. Eventually.

But it is a bad starting point.

If the client does not know the current cost of the work, the current quality of the work, the current conversion rate, the current review burden, or the current failure rate, then performance pricing is not clean. It is guesswork with a nicer jacket on.

You can already see the market moving towards this argument.

OpenAI published a useful scorecard this week. Ignore the inevitable vendor gloss for a second. The important shift is the language. It talks about useful work per dollar. Cost per successful task. Whether the result is ready to use, needs correction, or needs escalation.

That is much better than counting prompts, tokens or seats.

It is also where AI pricing starts to get serious. Because if an agent is going to do work for a business, the business needs to know more than whether the model is cheap. It needs to know what a successful task costs end to end.

Model spend. Tool calls. Retries. Failed runs. Review time. Corrections. Approval. Risk checks. The awkward senior-person-cleanup cost nobody puts in the deck.

That is the buyer-side scorecard.

The seller-side version came through in the Marketing School / SingleBrain episode from 16 July. Eric Siu was talking about managed revenue agents and the commercial ladder they are using: a paid pilot first, then a monthly base, then outcome participation once the client can actually model the saving or upside.

That sequence matters. Pilot. Base. Outcome. Not outcome first.

The reason is simple: on day one, the client often cannot tell you what the outcome is worth.

They may know they are wasting time. They may know their website under-converts. They may know their sales follow-up is patchy. They may know their paid media reporting is a mess. They may know customer enquiries are not being handled fast enough.

But knowing there is pain is not the same as knowing the commercial value of fixing it.

That is what the pilot is for.

A good AI-agent pilot is not a demo. It is not a theatre week where everyone gets excited and then nothing changes. It is a paid measurement exercise that proves whether a workflow is worth operating.

Pick one recurring piece of work. Define what accepted means. Run the AI-assisted version against the current version. Count the full cost. Count the review burden. Count how often the result is ready to use, how often it needs correction, and how often a human has to step in.

Then you have a baseline.

And once you have a baseline, the pricing conversation changes. Before the baseline, outcome pricing is mostly vibes. After the baseline, outcome pricing can be a proper deal.

Take a website revenue-agent loop.

The weak pitch is: “We’ll use AI to improve your website conversion and take a slice of the uplift.”

Sounds good until everyone starts arguing about attribution, seasonality, offer changes, traffic quality, sales follow-up, discounting, CRM hygiene, and whether the uplift came from the agent or from the client’s sales director finally replying to leads on the same day.

The better pitch is narrower.

Run a four-week paid pilot. One conversion metric. One source of truth. One weekly experiment. Human approval before anything goes live. Measure shipped tests, conversion change, lead quality and review time. If the loop produces useful evidence, move to a monthly operating retainer. If the value becomes obvious and attributable enough, then discuss upside.

Less exciting. Far less stupid.

Same with an AI tools assessment for an SMB owner.

The value is not that Claude can write a report. Everyone can get an AI tool to write a report now. The value is the bounded diagnosis: discovery call, pain extraction, tool recommendations, human QA, simple report, review call, then a clear implementation path.

You can price that as a paid audit because the accepted outcome is obvious: a client understands the first few fixes and trusts you enough to help implement them.

Outcome pricing too early would make no sense there. What are you taking a percentage of? Hours saved? Revenue created later? A better sleep score for the owner because the invoice chasing is finally under control?

Start with the audit. Create the scoreboard. Earn the next step.

This is especially true for managed campaign agents.

A client might want an agent that monitors the market, suggests angles, drafts assets, checks performance, builds follow-up tasks and keeps the campaign moving. Fine. That is a real operating layer, not just content generation.

But the first sale should not be “give us 10% of the revenue our AI creates”.

The first sale should be a pilot with a defined lane: which campaign or offer is in scope, which sources the agent can use, what it produces each week, who approves it, what counts as good enough, what gets published, what gets rejected, which metric matters, and where the result is logged.

Then the base retainer pays for running the loop properly. The outcome share comes later, if the loop is clean enough to support it.

This is where a lot of AI-service sellers will get themselves into trouble. They will use performance pricing as a trust shortcut.

“No risk, we only win if you win.”

Nice line.

But if the terms are vague, the risk has not disappeared. It has just moved into the future argument.

The agency says the AI created the uplift. The client says the market improved. The agency says the agent saved 40 hours. The client says those hours were not worth £200 each. The agency says the lead was qualified. Sales says it was rubbish. The agency says the content shipped. The client says it would have shipped anyway.

Congratulations. You have invented attribution hell with more automation.

The fix is not to avoid performance pricing forever. The fix is to earn it.

For most AI-agent work, I would use a four-step ladder.

Step one: paid diagnosis. Do not give away the thinking. Map the workflow, the current pain, the systems involved, the commercial objective, the quality bar and the risk boundary. This can be a small audit, a workshop or a short assessment. The point is that the client pays for clarity.

Step two: paid pilot. Build or run the smallest useful version of the loop. One workflow. One owner. One metric. One approval gate. Time-box it. Log the work. Count what was ready to use, what needed correction and what escalated to a human.

Step three: base operating retainer. If the pilot works, keep running the system. This is the boring money: weekly cadence, source updates, prompt and workflow maintenance, QA, reporting, issue handling, improvement, and commercial review. It is not glamorous. Good. It is where trust is built.

Step four: outcome upside. Only add performance economics when the scoreboard is stable enough. Everyone should know the baseline, the data source, the attribution rule, the review process, the exclusions and the calculation.

That is a much healthier deal. It also makes the agency or consultant look more serious.

Any fool can say “we’ll be paid on outcomes”. A serious operator can say: “Here’s the baseline we need before that makes sense. Here’s the pilot that creates it. Here’s the retainer that operates it. Here’s where upside becomes fair.”

That is a better sales conversation because it treats the client like an adult.

It also protects the seller.

AI work is still messy. Agents get stuck. Source quality varies. Client data is often worse than anyone admits. Approval chains are slow. The person who asked for automation might not own the system the agent needs to touch. The CRM has custom fields from 2018 that nobody understands. The website analytics are half broken. The sales team has a private spreadsheet because of course they do.

If you price on outcomes before finding all that, you are not being brave. You are underwriting chaos.

This is why the post-agency opportunity is not just “we use AI”. That is already boring.

The opportunity is to design measured operating loops that make AI safe enough, useful enough and accountable enough to buy.

Foundry’s AI Marketing OS Audit fits this nicely. The audit is not a pre-sales call with a nicer PDF. It should be the baseline machine.

What recurring commercial work matters? What does it cost now? What breaks? Where is quality lost? Which sources are trusted? Who approves? What is the first accepted outcome? What is the cost cap? What should never be automated?

Answer those and you have something worth building.

The AI Campaign Operating System then becomes the running layer: research, positioning, assets, publishing, follow-up, measurement and learning, with human taste and approval in the right places.

Then, if the client wants outcome pricing, you can have the conversation from solid ground. Not before.

The next wave of AI buyers will not be impressed by another agent demo. They have seen the magic trick. Some of them have already paid for it and found the invoice more reliable than the output.

They will want proof.

What work did it do? What did it cost? Was it accepted? What needed fixing? What escalated? What changed commercially? What would have happened anyway?

If you can answer those questions, outcome pricing becomes possible.

If you cannot, stick to pilots and retainers until you can.

There is nothing unambitious about that. It is the grown-up version of selling AI.

---

Pull quotes

• “If nobody knows the current cost of the work, nobody knows what the AI agent saved.”
• “Before the baseline, outcome pricing is mostly vibes. After the baseline, it can be a proper deal.”
• “Do not use performance pricing as a trust shortcut. It only moves the argument into month three.”
• “Any fool can say ‘we’ll be paid on outcomes’. A serious operator can explain the baseline needed before that makes sense.”
• “If you price on outcomes before finding the mess in the client’s systems, you are not being brave. You are underwriting chaos.”

Short LinkedIn / X version

Outcome pricing sounds like the honest way to sell AI agents.

“Only pay us when it works.”

Fine. But against what baseline?

Most businesses do not know the current cost of the work they want AI to improve. They do not know the review burden, failure rate, attribution rules, or what a successful task costs end to end.

So jumping straight to performance pricing is not brave. It is usually a future argument.

The better ladder: paid diagnosis, paid pilot, measured baseline, base retainer, then outcome upside once the scoreboard is clean enough.

OpenAI’s latest scorecard talks about useful work per dollar and cost per successful task. That is the buyer-side shift.

The seller-side shift is this: earn the outcome model before you sell it.

Notes and caveats

• The SingleBrain pricing ladder comes from the 2026-07-16 Marketing School episode note, based on local YouTube auto-captions. Treat the specific dollar examples as podcast-reported and source-check before using them in a published article.
• OpenAI’s scorecard is an official vendor source. Useful language, but it is not neutral research. Keep the point, avoid treating it as independent proof that OpenAI’s models always reduce cost.
• The article deliberately avoids claiming that outcome pricing is bad. The claim is narrower: outcome pricing is risky before baseline, attribution and acceptance rules are agreed.
• UK pricing equivalents are not specified. If Jason wants a Foundry offer page from this, convert the ladder into pounds and margin-test it separately.