The industry conversation on AI has quietly shifted from what it can do to what it costs. Trade press this week called it the AI buffet ending and portion control beginning. For claims and legal operators, this is a governance question, not just a budget one. Here is what portion control looks like in practice, and the three questions to ask before the next invoice.
For the last two years the AI conversation in claims and legal has been about capability. What can it draft, what can it check, what can it speed up. The price of finding out was low enough that most firms did not think about it too hard. Try a tool, see if it helps, expand if it does.
This week the conversation turned. Trade press described the shift as the AI buffet ending and portion control beginning, alongside separate warnings about an AI blind spot in governance as spend climbs. The point underneath both is the same. AI has stopped being a cheap experiment and started being a running cost, and a running cost with no owner is a governance gap, not just a budget line.
We work with claims and legal operators every week, and we think this is the right moment to get ahead of it. Not because AI spend is out of control at most firms, but because the moment to put controls on a cost is before it grows, not after.
Why "the meter is running" is the right image
The reason AI cost crept up on people is the pricing model. Traditional software is a licence. You pay a fixed sum, you use it as much as you like, the number does not move. You can forget about it between renewals.
Frontier AI does not work like that. The more capable models are metered. Every heavy query, every long document, every agent that goes off and does a multi-step task, all of it draws down. The bill scales with use. That is a genuinely different shape of cost, and it catches finance functions out because it behaves nothing like the licence line they are used to.
For a claims or legal operation this matters more than for most, because the work is document-heavy and volume-driven. Exactly the profile that runs up metered usage fastest. A tool that felt free in a pilot on ten files behaves very differently across ten thousand. Nobody did anything wrong. The meter was simply running the whole time, and nobody was watching it.
The second cost most firms are not counting
There is a second cost that does not appear on any invoice, and it is the one we would flag hardest for regulated operators.
When your people put client material into a general AI tool to get work done, the fee is not the only thing you are spending. You may also be spending know-how. Depending on the tool and the terms, the content going in can inform the model or sit somewhere you do not control. For a claims file or a legal matter, that is not an abstract risk. It is confidentiality, and in some cases privilege, leaving the building in exchange for a bit of speed.
So the honest cost of an ungoverned AI habit is two numbers, not one. The metered fee you can see, and the material you cannot get back once it has gone. Portion control is about both. It is not only "how much are we spending", it is "what are we spending, and are we comfortable with all of it".
What portion control looks like in practice
Portion control sounds like a budgeting exercise. It is really a governance one, and it comes down to three questions any operations or compliance lead can ask this quarter.
One. Does every AI tool in use have a named owner? Not a user, an owner. Someone accountable for what it costs, what it touches and whether it should still be in use. If a tool is running across a team and no single person owns it, that is the gap to close first. You cannot control a cost that belongs to everyone and no one.
Two. Is AI spend on a cost code you can actually see? Metered AI cost that lands in a general software or overhead line is invisible until it is large. Put it somewhere it can be watched month to month. You are not trying to minimise it, you are trying to see it, because a cost you can see is a cost you can make decisions about.
Three. Is there an audit trail for what the tools can reach? For each AI tool, someone should be able to say what data it can access and what it does with it. This is where the cost question and the confidentiality question meet. If nobody can answer it for a tool that touches client files, that is the finding, and it is a bigger one than the invoice.
None of these three is a technology project. They are operational decisions a firm can make in an afternoon. The value is not in the answers being difficult. It is in someone actually being accountable for them.
Why this sits on the compliance beat, not just finance
It would be easy to file all this under cost control and hand it to finance. We would push back on that.
The reason is that the two costs are linked. The tools that run up the metered bill fastest are the same tools people reach for to move quickly through client work, and moving quickly through client work is exactly where the confidentiality exposure lives. You cannot govern the spend without also seeing what the spend is touching. That makes this a compliance and operations question with a finance dimension, not a finance question with a compliance footnote.
The firms that handle this well will not be the ones who banned AI to control the cost, and they will not be the ones who left the meter running unwatched. They will be the ones who named the owners, made the spend visible, and knew what their tools could reach. That is portion control. It is not about eating less. It is about knowing what is on the plate.
Where CaseFlow Automation fits
We build compliance-first workflow platforms for claims and legal operators, and the design principle behind them is the answer to this whole question. The work runs on grounded, auditable outputs rather than open-ended queries into a general model. You can see what the tool did, what it drew on, and what it touched. The cost is predictable because the tool is built for a defined job, not billed by the appetite of whoever is prompting it.
That is the difference between AI as a metered free-for-all and AI as a governed part of the operation. If a firm's AI use currently has no named owner, no visible cost line and no clear answer on what the tools can reach, this quarter is a good quarter to fix that, and we are happy to talk it through with any claims or legal operator who would find a specialist steer useful.
CaseFlow Automation builds compliance-first workflow platforms for insurers, law firms and claims operators.
Frequently Asked Questions
- Why is AI becoming a governance issue and not just a cost?
- Because the two costs are linked. The tools that run up the metered bill fastest are the same ones people use to move quickly through client work, which is where confidentiality exposure sits. You cannot control the spend without also seeing what the spend is touching, so it belongs with compliance and operations, not finance alone.
- What does "portion control" on AI spend actually mean?
- Three things: every AI tool has a named owner accountable for it, AI spend sits on a cost code you can see month to month, and there is an audit trail of what each tool can reach and do. It is about visibility and accountability, not about using AI less.
- Why does metered AI cost catch firms out?
- Traditional software is a fixed licence you can forget between renewals. Frontier AI is metered, so the bill scales with use. For document-heavy claims and legal work, a tool that felt free in a small pilot can behave very differently at full volume.