• Welcome to Digital Expressway Consulting
  • +91 98180 30135
  • info@digitalExpressway.net

Why Only 2 in 10 Companies Make AI Agents Work, and the 4 Moves That Set Them Apart

← Industry Foresight & Innovation Hotbeds · AI agents

Why Only 2 in 10 Companies Make AI Agents Work, and the 4 Moves That Set Them Apart

One bank gives each of its AI agents an employee number, a login and a human manager, according to a Microsoft report on BNY. Another, Brazil’s Nubank, rehearsed its customer-support AI on more than 16,000 simulated customers before a single real customer saw it, according to a paper by its own team. A third, Singapore’s DBS, counts no value from AI unless the results beat a comparable group that did not use it, according to the research firm Forrester.

Meanwhile, in an independent research test, the best AI agent completed only 30% of ordinary office tasks on its own.

Same technology, very different results. Only about 2 in 10 organisations have scaled AI agents across the business, according to McKinsey’s 2026 survey. The difference is rarely the AI itself. It comes down to a handful of decisions about how the work is designed, checked and led, and every one of them can be learned.

What you will take away from this article

  • Four moves leading organisations use that your team can copy.
  • The four reasons most agent projects stall, so you can avoid them.
  • What the successful few do differently, side by side with everyone else.
  • A week-by-week plan to run your first safe agent project in 30 days.
  • The five habits that separate trained teams from those who stumble.

About a 12-minute read.

First, a quick definition. Most people first met AI as a chatbot: you ask a question, it writes an answer, and you decide what to do. An AI agent goes a step further. It is software, powered by an AI model, that plans a sequence of steps and carries them out across your systems: reading documents, looking up records, filling in forms, writing or changing code, and passing work to a person when a decision is needed. A chatbot answers; an agent acts, within limits that people set.

Four Moves the Winners Make

These four ideas come from organisations that are getting real results from AI agents. Each one is simple enough to adapt for your own team.

  • Give each AI agent an employee number and a manager. BNY, the US bank, according to a Microsoft WorkLab report (Microsoft is one of the bank’s technology suppliers), treats its nearly 140 AI agents as “digital employees”: each has its own login, a staff directory entry and a human manager who assigns and reviews its work. The rules you already use for people (access, review, joining and leaving) make the AI accountable too.
  • Rehearse on simulated customers before going live. Nubank, a large Brazilian digital bank, according to a research paper written by its own team, tested its card-support AI agents in more than 16,000 practice conversations with computer-generated customers, then ran a live comparison. The share of customers who solved their problem without a human rose by about 9 percentage points, with no drop in satisfaction.
  • Measure against a group that did not use AI. DBS, Singapore’s largest bank, according to the research firm Forrester, compares results from each AI service with similar cases handled without it, counting only the difference as value. It reported reaching its target of 1 billion Singapore dollars (about 770 million US dollars) of AI value in 2025.
  • Recruit volunteer champions in every team. Citi, according to the news site AI News drawing on Business Insider reporting, built an internal network of about 4,000 employees among its 182,000 staff, including volunteer “AI Champions” who act as the local go-to person for colleagues. More than 70% of staff now use the bank’s approved AI tools.

Why Most Agent Projects Stall

Agents fail in four predictable ways, and almost none of them are about the AI being “not clever enough”.

1. They are not yet reliable on messy, real work. In a 2025 research paper called TheAgentCompany, researchers built a simulated software company and gave AI agents everyday office tasks. The best agent completed only 30% of tasks fully on its own. In another test, τ-bench, agents handling customer requests in a simulated shop succeeded on all eight attempts at the same task less than a quarter of the time. An agent that works on Monday may not work the same way on Tuesday.

2. Most projects stall between the trial and real use. Surveys disagree on how widely agents are used, because they measure different things. PwC found 79% of 308 US executives saying agents are being adopted in their company, while McKinsey found only about 2 in 10 organisations scaling them across the business. A widely discussed MIT study of generative AI pilots in 2025, as reported by Fortune, found only about 5% delivering rapid revenue growth, though it was based on a limited set of interviews, a survey and public examples. The pattern is consistent: trying AI is common, getting measurable results is not.

3. Connecting to existing systems is the hardest part. In Anthropic’s survey of more than 500 technical leaders (December 2025), the top barriers were connecting agents to existing systems (46%), getting access to good-quality data (42%) and managing the change for people (39%). Note that this survey was run by an AI company. In India, EY found 78% of enterprises struggle with the same integration problem.

4. Security has no complete fix. An agent that reads emails and documents can be tricked by prompt injection: instructions hidden inside a document or email that the AI mistakes for genuine orders. OWASP, an open community that publishes widely used security guidance, says it is unclear whether fool-proof prevention exists. It lists “excessive agency”, giving an AI too many functions, permissions or freedom, as a top risk. The defence is to give agents the least access they need, and to require a person to approve any action that matters.

Case in point: Klarna, the payments company

  • What happened: In February 2024, Klarna announced that in its first month its AI assistant was handling two-thirds of all customer-service chats, work the company said was equal to 700 full-time staff.
  • What went wrong: By May 2025, according to Fortune’s report of a Bloomberg interview, Klarna’s chief executive said cost had been “a too predominant evaluation factor” in the decision, and the result was “lower quality” service.
  • What changed: He said Klarna would recruit more people so that customers can always reach a human if they want to.
  • The lesson: The technology worked. Judging it mainly on how much money it saved, rather than on the quality customers received, did not.

What the Successful Few Do Differently

The organisations that get real business results from AI are a small group, and they differ from the rest mainly in how they organise work and lead people. The consulting firm BCG puts it as a rule of thumb: about 10% of the value from AI comes from the AI models, 20% from data and technology, and 70% from changing how people and processes work.

In McKinsey’s 2026 survey, the “high performers” (about 6% of organisations, each earning at least 5% of its operating profit from AI) stood out in a few clear ways. BCG’s study of 1,250 companies found the same pattern among its most successful 5%.

What successful organisations do What it means in practice The evidence
Redesign the work, not just add AI Rethink a process from start to finish around what AI and people each do best About 3 in 4 high performers do this, against about 1 in 4 others; in McKinsey’s 2025 survey of 25 factors, it had the biggest link to profit
Measure against clear targets Agree in advance what success means in numbers, then track it Over 60% of BCG’s most successful companies track value rigorously, against 17% of the rest
Test before release Business experts check the AI’s work on real tasks before anyone relies on it According to OpenAI’s case study, Morgan Stanley had advisers grade answers against realistic scenarios and ran daily tests before scaling
Have visible leadership Senior leaders sponsor AI personally and set targets Nearly all of BCG’s most successful companies have engaged senior leaders, against 8% of the weakest
Give a business owner responsibility A manager in the business, not only a central tech team, owns each project and its results MIT’s 2025 study, as reported by Fortune, found successful organisations empowered line managers to lead adoption
Train most of their people Real training time, plus learning from colleagues The most successful companies expect to upskill over half of their staff, against about 1 in 5 elsewhere

These figures come from surveys, so they show what successful companies have in common rather than proving cause and effect.

Where Agents Already Work

Where are agents actually in use? In McKinsey’s State of AI survey of 1,719 organisations in 97 countries (August 2026), about 2 in 10 said they have reached the stage of scaling agents across their organisation. Among companies with revenue above $1 billion, 40% are scaling agents in at least one business function, against 22% of smaller companies. The functions where agents are most often in wide use are IT, knowledge management (finding and organising company information) and software engineering.

Three examples show the range. Each is labelled by how much we can rely on it.

  • Moving old software to new systems (Amazon, reported by the company, 2024). According to Amazon Web Services, Amazon used an AI coding agent to move tens of thousands of its own applications from older versions of the Java programming language to a newer one. The company says this saved the equivalent of more than 4,500 years of developer work and about $260 million a year.
  • Helping financial advisers find answers (Morgan Stanley, described by its AI supplier). The bank gave its advisers an AI assistant that answers questions from the firm’s own research and documents. According to OpenAI, which supplies the tool, before releasing it the bank had advisers and engineers grade its answers against realistic scenarios, and ran daily tests with a fixed set of sample questions; OpenAI also reports that more than 98% of advisor teams now use it.
  • Opening bank accounts faster (Incore Bank with Kyndryl, a trial, 2026). According to Kyndryl, the bank’s technology partner, a Swiss bank tested a team of AI agents that read customer documents and prepare the identity checks banks must run before opening an account. In this proof of concept, Kyndryl reports that the agents extracted data with up to 99% accuracy, and says the approach could cut onboarding from months to days. It is a test, not yet a live service.

India view. In EY India’s research (November 2025), 24% of leaders said they are already deploying AI agents, while 78% said connecting AI to their existing systems is a struggle. A Nasscom study of more than 100 global enterprises found 62% experimenting with agents, and 76% using their own IT department as the first place to try them.

How It Works

There are two ways to put AI to work on a multi-step task, and choosing between them is the first design decision.

  • A workflow follows a fixed path that people design in advance: step one, then step two, then step three. The AI does the reading or writing at each step, but the order never changes. Workflows are predictable and easy to test.
  • An agent decides its own steps as it goes, choosing which tool to use and what to do next based on what it finds. Agents handle messier, less predictable work, but are harder to test and control.

Anthropic’s guidance on building agents recommends starting with the simplest design that works and adding the agent’s freedom only where it clearly pays. Anthropic reports that the most successful implementations it has seen used simple, combinable building blocks rather than complex frameworks.

Every agent has four parts:

  1. The AI model, which reads, reasons and writes.
  2. Tools: the specific actions it is allowed to take, such as searching a knowledge base, reading a record or updating a ticket.
  3. Access to data: the documents and systems it can see.
  4. Guardrails and a human checkpoint: rules on what it may not do, and the points where a person must approve before it acts.

The example below is illustrative, not a real company: an IT service-desk agent handling a common ticket.

[DIAGRAM: Illustrative IT service-desk agent: the agent does the routine work; a person approves anything with real impact]

The agent reads the ticket and the employee’s device details, looks up the likely fix, and applies routine fixes itself. Anything with real impact, such as changing someone’s access rights, goes to an IT engineer for approval first. Every step is recorded so it can be checked later.

Value It Unlocks

The best independent evidence shows real productivity gains, largest for less experienced staff, but not automatic for everyone.

  • Customer support. In a study of 5,179 customer-support staff at one company, published by the US National Bureau of Economic Research, those given an AI assistant resolved about 14% more customer issues per hour. Newer and less experienced staff improved by 34%, because the AI passed on the know-how of the best performers. Highly experienced staff gained little.
  • Professional teamwork. In a controlled trial with 776 professionals at Procter & Gamble (2025), one person working with AI produced ideas as good as a two-person team without AI, and teams using AI were the most likely to produce the very best ideas.
  • Capability is rising quickly. Researchers at METR, an independent AI evaluation group, measure how long a task AI can complete on its own, half of the time. That length has roughly doubled every seven months since 2019. What an agent cannot do well today, it may do within a year or two.

The gains are uneven, and sometimes the opposite of what people expect. In an early-2025 study by METR, 16 experienced software developers working on their own projects took 19% longer with AI tools, even though they believed AI had made them about 20% faster. A larger follow-up in 2026 with 57 developers pointed towards a speed-up, but METR itself describes that evidence as weak. The lesson: measure the effect in your own work rather than assume it.

Where Claude Fits, and Where It Doesn’t

Claude, the AI model made by Anthropic, is well suited to the reading, reasoning and writing at the heart of most agent work, provided people stay in charge of decisions with real consequences.

Where it fits well

  • Working through long or many documents: contracts, policies, reports, tickets and case files, pulling out what matters and explaining its reasoning.
  • Multi-step tasks with tools: looking things up, drafting, checking against rules, and updating systems through approved connections.
  • Writing and changing code, including routine updates to older software, with engineers reviewing the result.
  • Drafting for people to finish: first versions of reports, emails, analyses and plans that a professional then checks and completes.

Where a person should stay in charge

  • Moving money or committing the organisation: payments, approvals, contracts and anything legally binding.
  • Decisions about people: hiring, performance and customer eligibility, where fairness and accountability matter.
  • Anything that cannot be undone: deleting records, sending external communications at scale, changing access rights.
  • Facts that must be exactly right: figures, legal points and regulatory statements should be checked against the original source, because any AI can state something wrong with confidence.

The principle is the same whichever AI you use: the AI drafts, gathers and proposes; an accountable person decides.

How to Start: A 30-Day Checklist

A team can run a safe, measurable first agent project in a month by copying what successful organisations do and avoiding what failed ones did.

  • [ ] Week 1: Name an owner and a sponsor. One manager in the business owns the result; one senior leader backs it.
  • [ ] Week 1: Pick one problem. Choose a single internal, high-volume, low-risk task, such as answering staff IT questions or preparing a routine report. Starting with your own operations, as 76% of enterprises in Nasscom’s study do with IT, keeps customers safe while you learn.
  • [ ] Week 1: Redesign the process on paper. Map each step and decide what the AI does, what a person does, and what is no longer needed.
  • [ ] Week 2: Set a baseline and targets. Measure how long the task takes and how often it goes wrong today, so you can prove any improvement.
  • [ ] Week 2: Build a test set. Collect 30 to 50 real examples with known good answers, and have your own experts grade the AI against them before anyone relies on it.
  • [ ] Week 3: Start simple. Build a fixed workflow first; add agent freedom only where it clearly helps.
  • [ ] Week 3: Limit access and add approval points. Give the AI only the systems it needs, and require a person to approve any action that matters.
  • [ ] Week 4: Train the people who will use it, to a common standard, so the whole team recognises the same mistakes and shares what works.
  • [ ] Week 4: Review honestly. Compare results with the baseline, record what did not work, and decide whether to expand, change or stop.

For managing risk, the US National Institute of Standards and Technology’s AI Risk Management Framework sets out a structure in four steps: govern (set rules and responsibilities), map (understand the context and risks), measure (test and track) and manage (act on what you find).

Why Trained Teams Succeed

Most failed AI projects repeat the same short-sighted habits. Experts call these anti-patterns: ways of working that look sensible at first but reliably lead to poor results. A trained team recognises them early and does something different.

The short-sighted habit What a trained team does instead
Bolting AI onto the old process, so the same report is produced the same way with AI writing one paragraph Redesigns the process around what AI and people each do best
Jumping straight to an autonomous agent when a simple chat or fixed workflow would do Starts with the simplest design and adds AI freedom only where it pays
Trusting a confident, polished answer, or checking it by simply asking again Checks against the original source or expert judgement, with more care where the stakes are higher
Giving vague instructions, then blaming the AI Writes a clear brief: what is needed, how to produce it, and what good looks like
Treating access to a system as permission to act Gives the AI the least access it needs and keeps a person approving actions that matter

Training makes a measurable difference. In BCG’s study of more than 10,600 workers, people with more than five hours of AI training were more likely to use AI regularly (79%) than those with less (67%). Yet in BCG’s 2026 follow-up, only 36% of workers felt properly trained. When a whole team learns the same methods, it also gains a common language for spotting these habits and agreeing what to fix.

Build the skills to lead AI work. The CCAO-F prep course from Digital Expressway Academy prepares professionals who run AI work without writing code (consultants, managers, analysts, operations and delivery leads) for Anthropic’s Claude Certified Associate exam. Every module is built around the habits above: redesigning workflows, choosing the simplest design, briefing Claude well, checking its output, and keeping the right controls in place. See the course

Bringing AI to your whole team? Talk to us about our Industry Advisory, Innovation and Pilot Build, and Mentoring and Skill Development services.

Sources

All sources were opened and checked in October 2026. Figures described as company-reported come from the organisation itself.

Digital Expressway Insights. Published November 2026; last reviewed October 2026. Not affiliated with or endorsed by Anthropic. Claude is a product of Anthropic. Examples marked “illustrative” are composites created to explain an idea and do not describe a real organisation.