Anthropic made an announcement in May that I think deserves more attention than another model launch. Together with Blackstone, Hellman & Friedman and Goldman Sachs, it is forming a new AI services company focused on helping mid-sized businesses put Claude into important operations.

Anthropic’s explanation is familiar to me: putting Claude into core operations requires hands-on engineering and deep familiarity with how a business runs. I have spent the best part of two decades around software adoption and value. Buying a capable product has never been the same as getting the benefit from it.

The frontier labs have spent years racing to make models more capable. Now some of the most interesting investment is going into helping companies actually use them.

In my experience, implementation can be bumpy, but we usually get the system live. The harder gap is often what follows: whether people actually change how they work and whether that change is worth anything.

Better models do not redesign businesses

There is a seductive logic to enterprise AI. The models keep getting better, so implementation should keep getting easier. At the task level, that is often true. A stronger model can require less prompting, handle more ambiguity and succeed on work that previously needed elaborate scaffolding.

But a business process is not a prompt. It has inputs, systems, permissions, exceptions, incentives, policies, owners, customers and consequences.

A model can be brilliant at analysing a document without knowing whether it is allowed to access that document. It can generate a persuasive recommendation without understanding who is accountable for acting on it. It can automate 90% of a workflow while making the remaining 10% more important. That is why capability alone does not deliver transformation.

The labs are telling us where the work is

Look at the direction of travel. Anthropic is backing a services organisation that will place applied AI engineers alongside customers.

OpenAI and PwC announced a collaboration in May around finance workflows, including planning, forecasting, reporting, procurement, treasury, tax and accounting close. OpenAI described its own finance organisation as “customer zero”, using real operations to validate workflows, controls and human-agent collaboration.

These are implementation stories. They are about getting inside the process. That should make business leaders slightly suspicious of AI strategies that begin and end with selecting a model.

The closer AI gets to valuable work, the less the problem looks like model selection and the more it looks like organisational design.

Start with expensive work

If I were looking for AI opportunity inside a business, I would not begin with a list of AI use cases. I would begin with the economics.

  • Where is expensive expert time being consumed?
  • Where are customers waiting?
  • Where does work repeatedly move between systems?
  • Where are decisions being made with incomplete information?
  • Where is capacity constrained by a small number of people?
  • Where does an error create disproportionate cost?

Those questions reveal problems worth solving. Only then should AI enter the conversation. Sometimes the answer will be an agent. Sometimes it will be conventional automation, better software, cleaner data or a process that should simply be removed.

An AI strategy that requires every problem to have an AI-shaped answer is not strategy. It is technology shopping.

Implementation creates proprietary learning

There is another reason implementation matters. The model is broadly shared. The learning generated by putting it into your organisation is not.

When a company deploys AI into a real workflow, it discovers things that cannot be learned from a product demo.

  • Which inputs are unreliable?
  • Where does the model become uncertain?
  • Which exceptions dominate the workload?
  • What context changes the answer?
  • What do employees trust?
  • Which approvals add safety and which merely add delay?
  • What metric actually captures value?

That knowledge accumulates. Over time, the organisation develops an implementation capability that competitors cannot instantly buy by subscribing to the same model.

A failed pilot can still be valuable if the business records what it learned and uses that in the next attempt. A pilot that exists only to demonstrate activity teaches very little.

Middle-market companies are interesting

Anthropic’s decision to focus the new services company on mid-sized businesses is also notable.

Large enterprises can assemble internal AI teams, hire specialist consultancies and absorb long transformation programmes. Smaller organisations often have less organisational complexity but also fewer specialist resources.

The middle market sits in an interesting position. These companies can have substantial workflows, valuable proprietary data and meaningful scale, while still being small enough to change quickly. For them, AI could create disproportionate leverage. But only if the implementation burden comes down. That is why services, tooling and reusable patterns around the models matter so much.

The opportunity is not limited by whether Claude or GPT can perform the task in isolation. It is limited by whether the business can put that capability safely into production.

The human expertise does not disappear

Recruitment is a useful example. Software can support the process, but it cannot compensate for a business that has not worked out what the role needs, which skills matter or how to assess them in an interview. If those decisions are poor, putting AI around them may just help us run a poor process faster.

There is a similar problem when the person buying a system and the person using it want different things. Management may value reporting while the employee experiences another place to enter data. Telling people to adopt it does not resolve that difference. They need to see how it helps them do the job.

The strongest implementations will often combine two forms of expertise. One understands what the technology can do. The other understands how the business actually works. You need both.

A brilliant AI engineer who does not understand a regulated finance process can automate the wrong thing beautifully. A domain expert who does not understand the technology may design around limitations that no longer exist. The interesting work happens where those perspectives meet.

This is one reason I am sceptical of the idea that AI transformation is primarily an IT project. Technology teams are essential, but the opportunity sits across operations, finance, sales, service, product and the commercial model itself. AI has to get close to the work.

Measure the change, not the deployment

Implementation also changes what should be measured. A model being available is not an outcome. An agent being deployed is not an outcome. Even high employee usage is not necessarily an outcome.

The useful measures sit downstream.

  • Time saved.
  • Capacity created.
  • Revenue influenced.
  • Cost removed.
  • Errors reduced.
  • Cycle time compressed.
  • Customer value improved.
  • New activity made economically possible.

That discipline protects companies from confusing adoption theatre with transformation. It also makes prioritisation easier. If a proposed AI project has no plausible path to meaningful value, perhaps it should not be a priority.

The last mile is becoming the strategy

The models will continue improving. That is good news. But each improvement makes the implementation question more, not less, important because a larger set of business activities becomes technically possible. The scarce resource becomes knowing which possibilities matter and how to realise them.

Anthropic’s May announcement is therefore more than a services story. It is a signal about where the market is maturing.

We spent the first phase of generative AI asking what the models could do. The next phase is about making those capabilities survive contact with actual organisations.

My hope is that AI can remove some of that work, rather than give people another interface to learn. I suspect we will need fewer screens for some activities. But that is a direction I am watching, not a reason to skip the work of understanding a process and the people in it.

Sources: Anthropic enterprise AI services company · OpenAI and PwC finance collaboration