AI's Proven, Its Value Isn't
By Professor Ashley Braganza
Cost cutting in customer service ended in 30% pay rises. That is not a projection. Ed Thompson, Senior Vice President for Market Strategy at Salesforce, has followed the first cohort of organisations that put AI agents live in customer service in spring 2025. By the autumn, those same organisations were paying the staff who remained 30% above the market rate to stop them walking. The business case promised headcount reduction. The outcome was a smaller, more skilled, more expensive workforce that the organisation could no longer afford to lose.
Ed spends his working days with Salesforce customers across Europe, and on this episode of The AI Adoption Podcast he brought the numbers: survey data, his own longitudinal contact with organisations, and a willingness to not be caught up with the hype.
The average agent count is real, and it describes nobody
The survey figure was 12 agents per organisation. Ed called it true and completely misleading in the same breath, and the second half matters more than the first. In a room of 20 customers, one may have 1,000 agents live. The next has 20. Around half may have none at all.
Twelve is arithmetic, not description.
The trajectory underneath is real. He tested it himself around customers through the spring: about two agents on average last September, four by December, twelve now, which is the hockey stick everyone was promised. But the aggregate hides the finding he considers the most interesting of the past year, that leading companies are getting further and further ahead. Tech and telecoms run a year ahead of other sectors, and inside those sectors individual companies run six to nine months ahead of their own competitors. The gap is widening, not closing.
For a board using an industry benchmark to calibrate ambition, this is a live problem: the benchmark averages an organisation that is sprinting with one that has not started. Matching it means matching nobody. The useful number is not the mean; it is the distance between your position and the organisation in your sector running nine months ahead of you.
Time saved is not money saved, and finance directors have noticed
Around three quarters of companies can evidence time savings. Around 10 per cent of finance directors say they can see a financial benefit. Ed takes both from surveys he says are multiple and easy to find, and his explanation for the gap is that the average employee saves 14 minutes, and that does not mean they are off the payroll. It means they go home on time and stop working at the weekend.
Generative AI arrived as a general productivity improvement, and Ed’s description of the result is not flattering: people writing emails, generating work slop, annoying colleagues with poorly written reports and badly created graphics. He concedes in the same breath that they are also being more productive in fact. Both halves are true. Only one of them reaches the accounts.
Agentic AI, which he dates at 18 months old, seems to move the financial needle in a way the productivity tools did not. Over 80 per cent of 2025's implementations were in customer service, and all of them aimed at cost reduction. On their own terms they worked; productivity improved sharply. Then the second-order effect arrived. Staff began handling multiple channels at once, and every question reaching a human was hard, because the easy ones went to an agent. The people left were highly skilled and difficult to replace. Thompson has talked to organisations now paying 30 per cent above market rate to hold on to the staff their business case had counted as savings.
I put the obvious objection to him directly. Business cases for technology have always been premised on a percentage of staff being removed, and AI is no different, so the pressure to lay off must arrive. His answer was yes, the pressure will be there, once the agents work and the organisation trusts them. He did not withdraw the pay rise evidence either. Both stand.
Token economics became a board question in four months
A token is the currency of a large language model. Models predict the next word in a sequence, sentences are chunked up, and tokens are attached to the chunks. The price behaviour matters more than the definition. Token prices fell for five or six years, by 99 point something per cent and then they went up, and only on frontier models. He named Anthropic and OpenAI. Lower end model pricing did not move.
The consequences are visible in organisations. Uber burnt through its entire 2026 budget in four months. Underneath all three sits the comparison that reframes every AI investment paper: running a traditional piece of workflow is 1,000 to 10,000 times cheaper than doing the same work by using tokens.
The distinction Ed draws is between building and operating. Burning tokens to build makes every sense in the world. Burning them to operate a process daily, at ten thousand times the cost of the workflow it replaced, does not. Salesforce is asking the same of itself, he says, looking at semantic layers and knowledge graphs to cut token usage in operation. The other lever is switching models. Z.ai's GLM 5.2 runs a couple of months behind the frontier at roughly a tenth of the cost per token, and 80 per cent of AI startups now use Chinese and open source models rather than leading edge ones. His conclusion is a governance point, not a technical one: do not bet on a single model.
The adoption failure is the inability to change, and leaders keep building around it
This podcast exists because adoption is where the value is won or lost, and Ed put the argument in a form that removes every technical excuse. Assume good enough data, clean and well documented processes, working triggers, guardrails in place, everything done properly. Does that mean employees accept the agent arriving in their job? Not necessarily. They push back, saying they will not work with that agent, or that it threatens them. The problem becomes an adoption problem at precisely the moment the technology stops being the problem.
The organisations getting this right are not running a hundred experiments. They narrow to three or four use cases, start small, and refuse to run proofs of concept in a sandbox. They use real customer, supplier and employee data, and they build iteratively with the people the agent will affect. Ed's explanation is a sentence any transformation leader should have on the wall: people do not like things being done to them. Offer to take away a tedious, repetitive, irritating task, and employees come on board.
Governance built for a slower technology now functions as a decision to do nothing, taken without anyone deciding it. Meanwhile the largest employment effect so far has landed where nobody was watching: middle management. Remove the mid-year review write-up, the compliance coaching, the travel sign-off, and a manager who spent half their time on admin is harder to justify. The result is delayering, one manager to ten instead of one to six. Ed does not think they are being fired; some retire, more return to sole contributor roles. That is a quieter outcome than the headlines, and a harder one to plan a career around.
The question the conversation could not settle
Ed rejects a jobs apocalypse, and his reason should worry his own customers more than the prediction reassures them: many companies are not competent enough to implement aggressively, because their data is poor, their processes are terrible and their tech debt is worse. He names the condition under which he would be wrong. Rapid implementation across an entire organisation would produce the apocalypse, and he has found almost no one able to do it. On his personal view it takes a decade, on the pattern of UK supermarket checkout jobs lost across ten years while those supermarkets ended up employing more people. So the honest question is not whether the jobs go. It is what the humans who stay are for, and his answer is that we go back to the 1970s and do far more face to face. If he is right, the operating model many organisations are building points the wrong way.
Listen to the full conversation
Website: https://theaiadoptionpodcast.com
Comments