Your AI service vendor sells outcomes. Who owns the number when it fails?

You've probably heard the pitch by now. Agents instead of people. Charged per resolved case, not per seat. Savings that make procurement smile.

CIO has a good piece on what is actually happening behind those slides. Venture-backed firms are buying support desks, finance-ops shops, and managed-service providers, then rebuilding them around AI. When renewal comes around, you often get sold "outcome pricing" as if that alone protects you.

If you run a Dublin SMB and you're looking at a managed AI employee (or anyone promising to run your back office with agents), that trend cuts both ways.

On one hand, buyers are finally being offered completed work instead of another tool to learn. On the other, if the supplier gets to decide what "done" means, you can end up with a green report and a service that is quietly falling over.

What is actually changing

Outcome pricing in business-process services is leaving the pilot stage. AI made it easier to put a result in a contract, so more deals are written that way. The big integrators are moving too, swapping time-and-materials for a share of savings. Waiting this out is not much of a plan.

Adoption is still early. Gartner's 2026 CIO survey put deployed AI agents at about 17% of organisations, with more than 60% expecting to be there within two years. Providers are moving faster than most buyers' vendor processes. That mismatch is where bad contracts get signed.

Same Gartner research expects more than 40% of agentic AI projects to be cancelled by end of 2027, mostly on cost, fuzzy value, and weak controls. And only a thin slice of companies calling themselves "agentic" are the real thing. A lot of it is agent washing: chatbots and old RPA with fresh branding.

So when someone quotes a resolution rate, treat it like an uptime claim on a sales deck. Believe it after you've seen it on work that looks like yours.

The story that should make you nervous

The CIO article opens with a public-sector support deal that looked brilliant on the vendor's resolution number. Then someone pulled the reopen data. Auto-closed tickets had been counted as wins. Users had stopped logging issues. The dashboard was green. The service was not.

That is the whole risk with agentic workflows sold as a managed service. Whoever defines "resolved" owns the commercial relationship.

Cheap tokens are not cheap labour

There's a related warning doing the rounds, and it matches what we see in delivery. a16z put it bluntly: vague agent jobs create expensive retry loops. The agent keeps calling itself to fix itself because nobody wrote a clean job. You end up spending tokens on spending tokens.

OpenAI's notes on measuring AI agent value land in the same place from the buyer side. Useful work per euro beats watching token prices.

And token spend is turning into its own budget headache. Most SMBs should not become token accountants. Buy a predictable ops result and let the provider worry about the stack underneath.

When a managed AI employee is set up properly, it looks like this:

• a narrow job
• a clear definition of done
• a spend or retry limit
• a human who gets the exceptions
• an outcome you can check without trusting their slide

Questions to ask before you sign

These are the CIO renewal questions, rewritten for Irish SMB owners. Use them with any AI-rebuilt vendor, or anyone pitching to run agentic workflows for you.

1. Who defines done, and on whose numbers?

"Resolved" should mean something your customers and team recognise. Take a baseline yourself before go-live. Tie the deal to metrics you own: reopen rate, time-to-resolution, backlog cleared, follow-ups completed.

If they push back on locking that down, believe them. That is the signal.

2. Real agentic system, or agent washing?

Ask for production results on work like yours, and the human-escalation rate sitting behind the headline number. A polished demo is not that.

3. Can you see what happened after the fact?

You want logs, clear data access, and an incident plan they have actually tested. In the EU that is not optional decoration. If an agent misroutes data or botches a dispute, people come looking for you.

4. Where does autonomy stop?

Write down what the managed AI employee can do alone, what needs a human, how the handoff carries context, and who is accountable when it acts. Put names on it before anything goes live.

5. What is the exit?

These platforms get sticky quickly. Lock data portability, who owns the knowledge base, and a way out while you still have leverage.

6. Are you buying capacity, or just a smaller invoice?

Sometimes the win is not cutting headcount. It is catching the work your old process was quietly dropping: unanswered tickets, stale follow-ups, leads that died in the inbox. Ask what that is worth before you only negotiate price.

7. Who pays for the tokens?

If the answer is "you, on a usage dashboard," you bought a tool with homework attached. If the answer is "we run the stack, you pay for the outcome," you bought a managed service.

Write the job before you flip the switch

Before the first loop goes on autopilot, fill in a short job card. Inbox triage and lead follow-up are good starters.

1. Trigger: what starts it?
2. Allowed actions: what can it do without asking?
3. Definition of done: what proves it finished?
4. Spend / retry limit: when does it stop trying?
5. Human exception: who gets it when stuck?

If you cannot fill that in, you are not ready. You are still designing the role. That is also how we scope Agentic Exp pilots: one costly coordination loop, one metric you own, thirty days of proof.

Where we sit in this

We're on the same side of the market shift CIO describes. Sell completed operational work. The difference is how we earn the right to charge for it.

We write the job before the demo. We run against your baseline, not just ours. Escalation and audit sit inside the retainer. Pricing is against labour value (executive tier €3,500-5,000/mo, value-based wedges for narrower SMB loops), and we absorb the delivery-stack economics. We're based in Dublin, with EU data boundaries treated as part of the product, not an afterthought.

You shouldn't need to wire ChatGPT Work to your calendar, babysit recurring prompts, and still own the failures just to get meeting prep and email drafts done. Those workflows are useful. The setup burden is exactly why managed delivery exists.

Bottom line

Outcome pricing is coming whether you like the pitch or not. It only helps if you own the definition of done. Vague jobs plus cheap tokens make expensive loops. A credible managed AI employee has a narrow role, limits, logs, a human brake, and an exit.

Start with one measurable workflow. Pilot it against your own baseline. Expand after that.

Want to pressure-test a supplier deck, or map your first loop? Book a free ops audit. We'll write the job card and the success metric before anyone switches anything on.