AI Adoption Is Levelling Off: Where Business Value Will Come From Next

AI adoption levelling off, illustrated by a rising bar chart that plateaus, with a path towards business value and discarded reports representing the shift from adopting more tools to achieving measurable results.

AI adoption is levelling off, and that changes the question boards should be asking. The issue is no longer whether the business uses AI. It is where the return will come from.

New figures suggest US business adoption of AI tools is nearing a plateau, and the AI infrastructure trade is under pressure. Ramp's 14 October index shows slowing growth in new customers and almost no growth in token spend. AI chip stocks lost over $1tn in value in the week of 12 October, and CoreWeave's debt unsettled investors. Workday launched Sana on 13 October, which shows major software vendors embedding agents in their own products.

Bain finds most enterprises are still stuck on narrow tool deployments, while those that treat AI as a full business change report 10% to 25% EBITDA growth. That gap points to where business value will come from next: bespoke work tied to specific workflows, not buying more licences.

This article is for CEOs, managing directors and boards reviewing their AI return on investment. It covers what the adoption and spending data is saying, why off-the-shelf AI tools stop adding value beyond a certain point, how to choose high-value workflows and measure the results, and a 90-day route from pilot to a measurable business outcome.

What the adoption and spending data is saying

Four sets of evidence landed within three weeks of each other. Read together, they say the same thing: the easy growth in business AI adoption is slowing, and the pressure is moving from buying AI to proving what it returns.

Ramp: growth in new customers has slowed and token spend is flat

The Ramp AI Index, published on 14 October by Ramp's economics lab, uses Ramp's spend data to track how American businesses use AI. In September, 55.9% of businesses on Ramp paid for AI tools, up 0.62 percentage points on August. The headline number is still rising, but Ramp's lead economist, Ara Kharazian, is clear about the trend: "the growth in new AI customers has slowed down significantly." In his words, "It feels like we're at the top of the S-curve, where the growth in the market will have to come from more intense adoption rather than more customers."

Spending tells the same story. Ramp recorded "two straight months of virtually no growth in token spend" in August and September, even though token volumes rose over the same period. Kharazian's view is that falling prices at the frontier are pulling the blended price per million tokens down. Spend is also concentrated. Ramp has previously reported that 80% of spend on OpenAI and Anthropic comes from 1% of firms, and even those firms spent $9,568 per employee per month in September, 3.1% below their July high of $9,873.

Two points matter for a board. These are US figures for businesses on Ramp's platform, so treat them as a signal of direction rather than a census. And flat spend with rising usage means AI is getting cheaper to use, not that businesses have stopped using it. The growth from new businesses signing up is slowing. What comes next depends on how deeply AI is used in the work itself.

Bain: most enterprises are still stuck at tool deployment

Bain's Technology Report 2026, published on 29 September, says that "as much as 90% of enterprises today remain focused on tool deployment and narrow use cases." Bain describes the result as micro productivity, with little impact on revenue or earnings. The leaders, a group Bain frames as roughly the top 10%, run AI as a full business transformation with C-suite sponsorship, and Bain says they are seeing 10% to 25% EBITDA growth.

Bain bases the 90% figure on its own experience with clients and does not publish the sample behind the EBITDA range, so read both as Bain's assessment rather than survey results. The direction is still useful: the gap between the two groups is not the tool they bought, but what they changed around it.

Markets: the AI infrastructure trade is under pressure

In the four trading days to 15 October, the 20 most valuable global AI and semiconductor stocks lost more than $1 trillion in market value, according to CNBC as reported by AI Weekly and SBS News. Reports pointed to concern that higher interest rates would raise the cost of financing AI data centres, to the scale of AI spending, and to growing Chinese competition in memory chips.

On 14 October, CoreWeave's shares fell about 15% as yields on its bonds neared 15%, on concern about AI borrowing. AI stocks have since rebounded, but CNBC reported on 19 October that CoreWeave's bond yields stayed near 15%.

For a business buying AI, the relevance is indirect but real. Investors are asking whether the money spent on AI infrastructure will be repaid by business value. Your board should ask the same of your own AI spend. We looked at the gap between AI infrastructure spending and revenue in our piece on the AI infrastructure boom.

Vendors: agents are moving into the software you already pay for

On 13 October, Workday launched Sana as the new Workday Home. Employees ask for what they need in natural language, and Sana coordinates Workday agents to complete the work, such as approving expenses, processing leave requests and resetting passwords, within each organisation's policies and permissions. The new home is available now. Sana Concierge, a new HR and IT service agent, is with early adopter customers and expected to be generally available in the first quarter of 2027. More than 11,500 organisations use Workday.

Sana shows major software vendors embedding agents in their own products. That is good news for buyers, because a capability that once needed a separate project now comes with the licence. It also means the same capability comes to every other customer of that vendor.

Why off-the-shelf AI tools stop adding value beyond a certain point

Off-the-shelf AI tools are good at making individuals faster. They stop adding value when the gains a business needs sit in how work moves between people and systems. A tool bought by everyone cannot give any one business an advantage, and a tool built for everyone does not know your processes.

Licences buy individual productivity, not a changed process

Bain is blunt about narrow tool deployment, saying it mostly just paves the goat trails, with little impact on revenue or earnings. Bain's evidence points the other way: "Companies that redesign workflows are more likely to generate new value from their AI investment" than those deploying tools to individual employees, and "The most important work lies in process redesign and modernizing the data and application environment."

Bain also reports that, among the companies it works with, "for every dollar spent on technology, four are spent on people and process." A licence covers the dollar. The return depends on the other four.

When every competitor has the same tool, it stops being an advantage

When agents arrive inside the software every competitor already uses, they raise the baseline for everyone. That makes them worth having, but not a source of advantage. The model underneath is no different. As Microsoft's Satya Nadella put it on the company's July 2026 earnings call, quoted in Bain's report, "any given model at any given time is swappable."

Generic tools stop at the edge of your own data and rules

A vendor's agents work on that vendor's data and processes. Workday, for example, says Sana works within each organisation's own policies and permissions. That is the right design for HR and finance records. But the workflows that set a business apart usually cross several systems and depend on rules that only that business uses, and no single vendor's agent covers them.

The trap

Adding more licences and expecting wider usage to turn into returns.

The fix

Choose the workflows where AI can change a number the board tracks, and build the integration, rules and review steps around them.

That is the case for bespoke work, and it rarely means building a model. In most cases it means connecting a commercial model to one specific workflow, with the data access, business rules and human review that workflow needs. Our work across multiple industries follows this pattern. We turned largely paper-based inspections into a digital workflow, and the system now drafts each Legionella risk assessment report for a compliance specialist to review and approve. The case study records significantly shorter report turnaround times.

Off-the-shelf tools still win for commodity work such as accounting and email, as we argued in our piece on off-the-shelf versus custom software. The point is not to replace them. It is to stop expecting them to deliver the gains that only a changed workflow can.

Choosing high-value workflows and measuring the results

A high-value workflow is one where AI can change a number the business already tracks. The choice of workflow matters more than the choice of tool, because the workflow sets the ceiling on the return.

How to choose a workflow

  • High volume and repetitionThe work happens often enough that small gains on each item add up.
  • A known cost baselineYou can say what each item costs today, in time or money.
  • One outcome the board recognisesCost per transaction, cycle time, error or exception rate, or revenue per customer.
  • Data the AI can reachThe inputs are digital, or can be made digital, and the systems involved can be connected.
  • A named ownerOne person is accountable for the workflow and its result.
  • Risk you can controlA person can review outputs wherever a mistake would be costly.

Our piece on AI document processing applies these tests to invoices and contracts, using four numbers: volume, cost per document, exception rate and cycle time.

How to measure the results

Measure the business outcome, not the activity. Ramp's data is a reminder that spend and usage can move independently: token volumes rose in August and September while token spend stayed flat. Neither number says whether the business got anything for its money.

Activity measures compared with outcome measures for AI return on investment
Activity measure Outcome measure to report instead
Licences or seats in use Cost per transaction in the chosen workflow
Prompts, queries or tokens Cycle time from start to finish
Hours saved, as reported by users Error, rework or exception rate
Number of pilots launched Change in cost, revenue or margin attributable to the workflow
  • Record the baseline before you build. Without it, there is nothing to compare the result with.
  • Compare like with like. Use the same period last year, or a team still working the old way.
  • Count the full cost. Include the people and process work, which Bain's experience suggests outweighs the technology spend by four to one.

A 90-day route from pilot to a measurable business outcome

Ninety days is long enough to change one workflow and measure the result, and short enough that a pilot cannot drift. The plan below is our framework, and the dates are a guide rather than a rule.

1Days 1 to 15: choose one workflow and record the baseline

Apply the tests above and pick one workflow. Agree the single outcome measure that will decide success, and record its current value from existing records. Name an owner for the workflow and a sponsor at board level.

2Days 16 to 45: build a narrow version inside the real workflow

Connect the AI to the systems the work already runs on, rather than building a separate tool people have to visit. Do one task end to end. Add a human review step where mistakes would be costly, and log every output so the results can be checked.

3Days 46 to 75: run it on live work and measure

Run the pilot on real volume. Track the outcome measure every week against the baseline, and record every exception and every time a person overrides the output. Adjust the rules, prompts and integrations as the evidence comes in.

4Days 76 to 90: decide with the numbers

Compare the result with the baseline and the full cost, then scale it, change it or stop it. If you scale, plan the process changes, training and ownership that make the workflow stick.

A pilot without a baseline cannot succeed, because there is nothing to measure it against.

Most pilots that stall do so between the demonstration and the workflow, as we set out in our pieces on why AI implementations fail and moving AI from proof of concept to production.

Where business value will come from next

The adoption curve is flattening, AI is getting cheaper to use, and the major software vendors are building agents into products every competitor can buy. Each of those trends makes generic AI more available and less distinctive.

The value that remains is in the work only your business does: the workflows that cross your own systems, follow your own rules and feed numbers your board already tracks. That is where businesses will find EBITDA growth, and it is where the next gains will come from, through bespoke work tied to specific workflows rather than more licences.

Is your AI spend tied to a measurable outcome?

If your AI review shows rising licence costs but no number the board recognises, start with one workflow, one baseline and 90 days.

At Imobisoft, we help businesses choose where AI will pay through our AI strategy and consulting work, and we design and build AI solutions around the systems and data they already have.

Talk to us about your AI workflows

Frequently asked questions

Is AI adoption levelling off?

Among US businesses, it appears to be. Ramp's October 2026 AI Index found 55.9% of businesses on its platform paying for AI tools, but growth in new AI customers has slowed significantly, and Ramp's economist describes the market as near the top of the S-curve.

What did Ramp's October 2026 AI Index find?

Business AI adoption rose to 55.9% in September, up 0.62 percentage points, while token spend saw two straight months of virtually no growth. Token volumes rose, so Ramp's economist puts the flat spend down mainly to falling prices.

What does Bain say about AI and EBITDA?

Bain's Technology Report 2026 says leaders running AI as a full business transformation are seeing 10% to 25% EBITDA growth, while as much as 90% of enterprises remain focused on tool deployment and narrow use cases.

Why do off-the-shelf AI tools stop adding value?

They make individuals faster but do not change how work flows. Every competitor can buy the same tool, and generic tools do not know your data or rules. The next gains come from workflows built around your business.

How do you measure the ROI of AI?

Record a baseline before you build, then measure a business outcome such as cost per transaction, cycle time or error rate. Count the full cost, including people and process, and report in terms that reach EBITDA.

How long does it take to get from an AI pilot to a measurable outcome?

About 90 days for one workflow: 15 days to choose it and record the baseline, 30 days to build a narrow version, 30 days on live work, and 15 days to decide with the numbers.

Sources

Have a project in mind? Let’s get to work.

Let’s chat about how we can help you. Fill in the details and we’ll get back to you as soon we can.