Agents turn employees into software producers: who manages, governs and pays for this growing agent estate?

McKinsey has 25,000 AI agents to augment its workforce and Microsoft's Digital business alone has 500,000. Starbucks, meanwhile, is building alternatives to software it currently buys from Microsoft and IBM, and its enterprise technology team targets around a $10 million reduction in software spend this year.

If you spend enough time reading about AI, it can feel as though everyone is now building their own software.

“Everyone is now empowered to build, but I think the challenge is also that everyone is now empowered to build,” says Jay Raval, former co-founder of group travel app Troopa and GTM Lead at Immutable.

What is interesting is that organisations are now empowered to build at much smaller scales. The result is a proliferation of agents that they now must also manage, govern, and balance against market uncertainty.

Granted, what we call agent varies widely from scheduled tasks to actual autonomous, self-looping LLM instances. Vendors are partially responsible for this conflation (read: confusion) by using ‘agents’ as shorthand for anything beyond a user chat.

Instead I find it helpful to think of agents as a spectrum of internal micro-ware, solving problems that are too specific to justify buying an entire product, too minor to justify a traditional software project, but irritating or frequent enough that solving them yourself has real value.

In a market this young, these smaller and more reversible AI bets are perhaps rational because no one yet knows where the technology will settle. But organisations that only make small bets can easily become very good at experimentation without ever building distinctive capability.

The challenge, then, is that every bet, however small, creates an ownership, maintenance and governance obligation for as long as it remains operational. Deciding which of them deserves greater commitment only gets harder the more of them you make.

How organisations are making smaller technology bets with agents

Ciarán Hennessy, Technology Program Director at TAL and fractional CTO, is seeing this play out less as a choice between building a software product or buying one, and more as organisations building their own smaller workflows or agents on top of AI infrastructure. There is now a category of builds that sit in an expanding middle between build and buy.

One of the more successful enterprise AI use cases Ciarán has seen is an onboarding agent that helps new employees find the information they need and raises the relevant tickets to get them set up. Onboarding is not core to an organisation’s operations, but it’s exactly the kind of repetitive, bounded problem that AI can handle well and, most importantly, it is now cheap enough to fix.

Micro-ware applications give organisations a way to act on AI and see what works and where, while keeping individual technology bets relatively small and reversible in an uncertain market.

Tom Bowden, Managing Director of Solsbury Advisory confirms that he’s seeing his clients take a trial and error approach. They are open to seeing where AI delivers value “while also waiting and watching for the next evolution.”

What we build on is still being built

Reversibility remains valuable both because individual bets are currently cheaper, and also because the organisation can hedge against market and timing risk.

“We’re in chapter one of a hundred chapter book” - Tom Bowden

Tom compares today’s enterprise AI market to the DOS era of computing. We have a powerful underlying technology, without the mature application layer that eventually made computing accessible and useful to everyone.

In his view, a much larger wave of productisation is still ahead. He is already seeing products begin to “leap over” organisations’ own thinking about how they might solve a problem, with vendors turning up to say, “we’ve made it already, you can just buy it.” In some cases, the client had not yet realised the problem could be addressed with AI.

Equally, a team may build a useful layer today only for Microsoft, Salesforce or a frontier AI lab to release the same capability later.

Cheap to build does not mean cheap to own

There is a trade-off in being endlessly reversible and flexible, however. If most AI investment is a small build, organisations can accumulate thousands of useful agents that increase cost without ever producing a step change in capability or advantage.

It’s the same problem that everybody had around productivity applications and application sprawl - Ciarán Hennessy

And now agent sprawl is going to start translating into higher costs.

The market is increasingly moving from a subscription model to a consumption model. Costs are starting to add up as more users and teams within an organisation move toward building out their own autonomous agents or chains of agents. On a per token basis usage costs can seem trivial, but as more agents run end-to-end workflows, they consume more of an organisation’s available AI token usage. It is not uncommon to hear of employees accidentally running up costs in the hundreds of thousands or even millions.

At 25,000 little bits of micro-ware, each one needs to create enough value to justify shifting costing models, maintenance, and governance overhead. Small reversible bets come with running costs that are already causing some organisations to cancel licences.

Every piece of micro-ware needs an owner

Agents may be getting smaller and easier to build, but the organisational machinery required to keep track of, maintain and govern them also grows as the agent estate expands.

The size of the agent or workflow plays a role in its relative maintenance effort, with smaller more compact builds that are built and owned by the same person being easier to maintain than larger ones.

“The maintenance for the kinds of things I’m building is not ultra heavy... an API might break and I just have to fix the API. You can build things really quickly, use them in small volume use cases, and it’s not as difficult to maintain as people generally thinks it is.” - Jay Raval

At a 1-3 person scale updating instructions when the model changes, re-running evaluations, fixing APIs etc. isn’t hard, but at a larger scale, even within teams of more than 3-4 people, it becomes someone’s job to maintain agents.

At the other end of the market, it’s not just maintenance but governance of agents that is pushing a rethink of infrastructure, capabilities, and processes. Microsoft has had to build an internal central registry and governance platform to track who owns each of its agents, where it came from, how it is being used, and most importantly what to do with it when the person that built it leaves the company.

This same control plane is now being rolled out to enterprise clients who are themselves struggling to manage their agent estate across divisions and teams. From what I’m seeing in the market, most IT teams have started to restrict or put a stop to agent building just to be able to catch up in their accounting.

What this means for ongoing maintenance and governance

Not everybody needs or wants to make outsized bet that companies like Starbucks are making, the outcome of which won’t be known for a few years. Neither do they want to buy lots of specialised software from startups that could or could not be around next year.

Building and testing micro-ware on top of existing infrastructure like AI lab subscriptions, Salesforce, or Microsoft environments is rational and reversible. But reversibility solves one problem while creating another — one of cost, maintenance and governance.

Self hosted infrastructure on open models is an alternative, but as of now the maintenance and governance burden is greater for self hosting, even if the cost is not.

What is becoming clear, is that we do not yet have a settled way of defining and therefore governing micro-ware agents. If every employee can build five or ten useful agents, at what point does a personal workflow become an enterprise asset? Who owns it when other teams start relying on it? Whose P&L pays for usage, who checks that it still works, and who decides when it should be retired?

The next phase of enterprise AI may not be large AI systems, but an ecosystem of personal, team and enterprise agents sitting on top of shared infrastructure. If that is where we are headed, governance shifts from managing software to managing entire agent estates.