TL;DR
AI is automating the interface layer that professional services firms have billed for, and the same shape is visible in accountancy, legal and marketing. This list sets out eight places the value moves, ordered by how fast each shifts. Together they show that AI disruption in professional services is a repricing problem before it is a tooling problem, and that a firm’s evolution rate decides the outcome.
On 26 August 2026, Salesforce and Anthropic announced Claudeforce. Its first product puts Salesforce inside Claude with thirty-seven prebuilt sales skills, live data and central authentication, and the stated aim that sellers will not need to open the application. It reached pilot customers the same day, with open beta stated for September 2026 and live demonstrations scheduled for Dreamforce between 15 and 17 September. Marc Benioff summarised the bet on the earnings call in five words: “the UI is the AI.”
Underneath the announcement sits AIforce, which exposes Salesforce data and workflows to any agent through MCP servers, APIs and command line tools, built on the Headless 360 architecture introduced at TDX in April 2026. The product is new. The plumbing has been shipping for months.
Anyone in the Microsoft channel who read all that with a degree of comfort should look at what Microsoft has already shipped. The ERP MCP server reached general availability on 27 January 2026. A Sales MCP server followed in July. Microsoft announced general availability of the Customer Experience MCP Server for Service on 30 July 2026 with more than ninety tools. A Commerce MCP server arrived for retail, workforce engagement tools were added on 3 September 2026, and Business Central ships agents that read documents and prepare transactions before a person opens a record. The original static ERP MCP server retires on 1 October 2026, replaced by a dynamic successor, which is worth holding on to for anyone still treating this as a single product decision.
The interface layer is being removed on both platforms. The question for any firm that implements them is which of its billable hours were attached to it.
That question is not confined to software partners. An accountancy practice billing for bookkeeping and reconciliation, an agency billing for first-draft copy, a law firm billing for document review and a recruiter billing for CV screening are all looking at the same structure. Value that sat in a layer of human effort compensating for the limits of software.
This list is ordered by how quickly each shift is likely to arrive, fastest first. That ordering is the useful part, because it tells you roughly how much time you have for each. The first four are erosions. The next three are places value accumulates. The last one complicates the picture, and it is worth reading before acting on the rest.
1. Configuration hours attached to screens
Form layouts, view design, dashboards, model-driven app polish, bespoke reporting. On a mid-market implementation this is a substantial share of a functional consultant’s time, and it is the first thing a generated interface makes redundant. When the agent creates a view around the task, nobody is paying for the view.
This one is already moving. It does not require the technology to mature much further, because generating a usable interface for a specific question is well within what current tooling does reliably.
The uncomfortable detail is that this work was often the most predictable and most profitable part of a delivery estimate. It scaled without senior people. That is precisely why it is exposed.
2. Training and adoption built around navigation
If a user works in a conversational surface, then teaching people where to click stops being a product. So do the follow-up sessions that were quietly holding adoption numbers up.
What survives is different in kind. Someone still has to teach a team how to judge whether an agent’s answer is trustworthy, when to escalate, and what the agent is not permitted to decide. That is a genuine training need and most firms do not currently have a curriculum for it.
The reason this moves quickly is that it depends on the first shift rather than on any separate technology change. Navigation training disappears when navigation disappears.
3. Entry-level build and light automation work
Natural language agent design lowers the barrier for a competent finance manager to automate a process without calling anybody. Not everything, and often not well, but enough to take the bottom off the market.
This matters more than the revenue it removes, because entry-level build work is how services firms train people. A practice that loses its simple work loses the route by which juniors become seniors. The margin effect is visible within a year. The capability effect appears three years later, when there is nobody ready to run the complex work that survived.
4. Resale margin, as spend moves to consumption
Consumption pricing routes money to the vendor. Agent tooling is metered per call. The firm ends up further from the revenue and closer to being the party that explains the invoice.
This shift is slower because contracts and renewal cycles slow it down, but its direction is not in doubt. The firms that notice early tend to start charging for consumption modelling and cost control, which is a real service and a small one compared with what it replaces.
5. Permission and access design
Here the value starts accumulating rather than eroding.
An agent inherits the permissions of the person running it. Both vendors have designed it that way. So every over-generous role and every unreversed elevation surfaces the first time somebody asks a plain question. This is chargeable, defensible work with a clear risk case attached, and most customers have been deferring it since go-live.
It grows more slowly than the erosions above, because customers only fund it after the first uncomfortable incident or the first serious pilot. Firms that can demonstrate the exposure before the incident will get there earlier.
6. Schema semantics and recorded meaning
This is the shift worth building a service line around.
Field descriptions, table descriptions, option set labelling, deprecation marking. In a screen-driven system this was documentation, which is why it was always first out of scope. In an agent-driven system it is read at runtime and directly determines answer quality.
Every environment that has been running for a few years contains fields with generated names, duplicate columns where one is dead, and option sets whose meanings live in a spreadsheet on somebody’s desktop. The customer’s team works around all of it from memory. That memory does not transfer to an agent, and rebuilding it as machine-readable context is a project in its own right.
The reason this sits sixth rather than first is that customers do not yet know they need it. Demand arrives after a pilot returns a confident wrong answer.
7. Testing behaviour that is right most of the time
Most delivery teams know how to test a workflow that either fires or does not. Very few have a method for evaluating a system that is right most of the time and confidently wrong for the rest.
This will surprise people. It requires evaluation sets, acceptance thresholds, sampling regimes, regression testing against changing models, and a conversation with the customer about what error rate is acceptable in which process. None of that is currently standard practice in mid-market delivery, and none of it is billed the way testing has been billed.
It is last among the growth items because it needs a skill the market has not built yet. That is also why it will price well for the firms that build it first.
8. The work that does not move
Vendor timelines are not delivery timelines, and this item exists to keep the other seven honest.
Complex and regulated processes will keep deterministic screens for years. Where a decision must be auditable, repeatable and identical across a hundred users, a generated interface is a liability rather than a saving. Somebody still has to build those, integrate them, and prove they work.
The claim that thousands of clicks will disappear is marketing. Something closer to the truth is that the discretionary, exploratory, question-shaped part of system use moves to agents, and the controlled, evidential part stays where it is. That is a large change and it is not a total one.
There is a brake on the vendor side too. Reporting at the end of August noted that some storage and access controls for regulated buyers remain unfinished, which means the customers with the strictest data handling obligations are among the last able to take agent access at all. For a firm with a regulated client base, that is a longer runway than the announcements suggest and a reason to price the deterministic work with confidence rather than apology.
The risk in overreacting is real. A firm that abandons its delivery capability in order to reposition as an advisory practice may find it has given away the thing that made its advice credible.
A composite example
The following example is composite. It is drawn from patterns across mid-market implementation firms rather than from one named business.
A partner of about forty people, most revenue from mid-market implementations, billing configuration days at a standard rate against a signed-off requirements document. The model has worked for a decade. It produces predictable margin and it scales delivery without scaling senior people.
The leadership team responds to the vendor announcements the way the channel usually responds to product change. Certifications, an internal enablement plan, a capability slide, a webinar. All of it sensible. None of it touches the pricing model, which is where the automation is landing.
Eighteen months later the certifications are in place and the average project is smaller, because the configuration days that used to pad the estimate are no longer defensible in a competitive bid. The firm is more capable and less profitable. The response to that is usually a rate cut, which accelerates the problem it is trying to solve.
The firms that hold are the ones that changed what they sold while they still had the margin to fund the change.
What the eight add up to
Read together, the list makes a single point. The technical change is happening at vendor speed, which is measured in months. The commercial change has to happen at firm speed, which is usually measured in years, because it involves pricing, sales skills, delivery method, hiring, incentives and how partners are compensated.
Vendor speed has just been made explicit, and in a way that lands directly on the channel. On 25 August 2026, the day before Claudeforce, Microsoft announced that release waves are finished. There is no Release Wave 2 for 2026. New capabilities are published to the AI at Work roadmap as they are committed, and Release Planner is withdrawn on 15 November 2026.
For a partner that matters twice over. The twice-yearly wave was a planning artefact for customers and a selling rhythm for the channel. It was the reason to run a roadmap session, the hook for an enablement quarter, and a reliable moment when a customer would agree to look at what was coming. That rhythm has gone, replaced by a continuous feed the customer can read for themselves and filter to their own environment. Any partner whose customer engagement plan assumed two natural conversations a year now needs a different reason to be in the room.
Edge151 calls the distance between those two speeds the repricing gap. It is the difference between how fast a firm’s work is being automated and how fast that firm can change what it charges for. Every industry facing AI disruption has one. The size of the gap, rather than the size of the threat, is what determines the outcome.
Closing that gap is a capability problem. Edge151 defines Business Evolution as the deliberate, continuous building of organisational capability at the rate a business’s ambition requires, and in this situation the rate is set externally. The vendors are setting it and they are not going to slow down to suit anyone’s planning cycle.
The instinct in most services businesses will be to treat this as a product transition, because that is what services businesses are good at. Learn the tooling, get the certifications, update the capability slide. All of that is necessary. None of it addresses the pricing question, which is the one that decides whether the next three years are comfortable.
A programme will not fix it either, for the reason set out in why transformation programmes decay. A programme finishes. The vendor release cycle does not.
Conclusion
Exposure in this shift tracks two things. How tightly a firm’s price was attached to the task being automated, and how slowly that firm can change what it sells. Technical sophistication barely features.
That combination is worth measuring honestly. A firm can be highly capable technically and still have a wide repricing gap, because the constraint sits in commercial authority, sales capability and leadership appetite rather than in skills.
The firms that start the commercial conversation now will have more room than the ones that wait for a bad quarter to force it. That claim rests on how quickly organisations can change what they are rather than on any forecast about AI, which is what Business Evolution means in practice.
If you are advising customers on how quickly they can adapt, apply the same test to yourself. Edge151 runs a Business Evolution Assessment that measures a firm’s rate of capability change across eight components, including the commercial ones that usually bind first. The guides set out how to do that.
Alastair Jupp writes on organisational capability and the practical adoption of AI. The argument here is developed at length in Workflows, Decisions, Discipline: The Operating System Behind Every High-Performing Business, published by Edge151 later this year. Details and publication updates.
The most exposed work is the effort that existed to compensate for what software could not do: interface configuration, navigation training, first-draft production, document review, reconciliation and screening. Exposure follows the compensating layer rather than the sector, which is why accountancy, legal services, marketing and technology implementation are all seeing the same shape.
Total hours for a given scope fall, and scope tends to expand because more processes become worth automating. Whether that nets out positively depends almost entirely on whether the firm can reprice. A practice billing configuration days loses, because the days are the thing being removed.
The repricing gap is an Edge151 term for the distance between the speed at which a firm’s work is being automated and the speed at which the firm can change what it charges for. It is a capability measure rather than a market measure, and two firms facing identical disruption can have very different gaps.
Four are already visible: permission and access design, schema semantics and recorded meaning, evaluation and testing of non-deterministic behaviour, and consumption cost modelling. Each is chargeable. None of them is billed the way configuration was billed, which is the difficulty.
Rate cuts accelerate the problem. If the underlying issue is that the billable task is being automated, a lower price on the same task shortens the runway without changing the position. Repricing means changing what is sold, not what it costs.
Faster than a planning cycle allows, which is the difficulty. Salesforce moved from announcement to pilot access in a day and has open beta scheduled within a month. Microsoft shipped three separate MCP surfaces in about seven months. A firm that reviews its commercial model annually is reviewing it at roughly a quarter of the speed of the change.
Previous shifts changed what the firm implemented. This one changes what the firm bills for, because the automated part and the billed part are unusually close to the same thing. That makes it a commercial event dressed as a technical one, and firms tend to be better equipped for the second than the first.
Discover more from Edge151
Subscribe to get the latest posts sent to your email.