TL;DR
- MIT's NANDA initiative found 95% of enterprise generative AI pilots delivered no measurable profit-and-loss impact, against an estimated $30 to $40 billion invested (The GenAI Divide: State of AI in Business 2025).
- That figure is widely quoted as "95% of AI projects fail". It does not say that. No measurable impact is largely a consequence of pilots never recording a baseline, so there was nothing to measure against.
- The same report found buying agent tools from vendors succeeded roughly twice as often as building them internally. That argues against most of what this industry sells, including parts of what Euracle sells.
- MIT also found over 90% of organisations had employees using personal AI tools for work while only around 40% had licensed tools officially. Your adoption signal already exists and is unmanaged.
- The highest-return two weeks of any agent project happen before the build, counting what the process costs today in hours, errors and cycle time.
Before reading further, check you actually need an agent. Our guide to AI agent vs chatbot covers the sorting question, and for a large share of business processes the answer is a much cheaper piece of software.
What are AI agents and how do businesses use them?
AI agents for business start from one idea: software given a goal rather than a script. You supply the outcome you want and the tools it may use, and it works out the sequence.
Three plain-English terms that make the rest of this readable:
- A large language model, or LLM, is the kind of artificial intelligence that reads and writes text. It is the part that does the thinking. On its own it cannot do anything except produce words.
- Tool use means giving that model permission to call other software: read a database, send an email, create a ticket. This is what turns a text generator into something that acts.
- Autonomy is how much of the sequence the software decides for itself. More autonomy means more capability and more ways to be surprised.
The practical difference from ordinary automation is that an automation runs the steps you wrote, in that order, every time. An agent picks its own route, so two runs of the same task can look different. That flexibility is the entire value and the entire risk.
Businesses use agents where the input varies enough that a fixed sequence breaks: incoming emails, support tickets, documents, research questions, exception handling. They should not use them where the input is uniform, because a fixed sequence is cheaper and more predictable.
Why do 95% of AI pilots show no return?
Mostly because nobody wrote down what the process cost before the pilot started. This is the central problem with AI agents for business today, and it is not a technical one. This is the single most important thing on this page and it is routinely reported as something else.
MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025, based on around 150 executive interviews, a survey of roughly 350 employees, and analysis of 300 public AI deployments. It found that 95% of enterprise generative AI pilots produced no measurable profit-and-loss impact, against an estimated $30 to $40 billion of investment.
95%: the share of enterprise generative AI pilots delivering no measurable profit-and-loss impact (MIT NANDA, The GenAI Divide: State of AI in Business 2025)
Read the wording carefully. The finding is "no measurable impact", not "the pilot did not work". Those are different claims and the gap between them is a spreadsheet.
If a company deploys an agent to handle support triage and never recorded how long triage took beforehand, how many tickets were misrouted, or what that cost, then at review time there is no before to compare the after against. The pilot may have improved things considerably. It will still be reported as delivering no measurable impact, because nothing was measured.
MIT's own diagnosis points the same way. The researchers attributed failure to what they called a learning gap in how organisations integrate these tools, explicitly not to model quality. The models were good enough. The deployments were not instrumented.
Two further findings from the same report are more useful than the headline:
- Buying beat building. Purchasing tools from specialist vendors succeeded roughly twice as often as building internally. More on this below, because it argues against a good deal of what agencies sell.
- Shadow use is already widespread. Over 90% of organisations reported employees using personal AI tools for work, while only around 40% had officially licensed tools. Your staff have already told you which tasks are worth automating. Nobody asked them.
So the correct response to the 95% statistic is not to avoid agents. It is to spend two weeks measuring before spending three months building.
What does an AI agent actually do in a business?
AI agents for business handle work where the input arrives in an unpredictable shape and the output can be checked. Here is what that looks like across four functions, at the level of a specific task rather than a department.
| Task an agent handles well | Why it fits | |
|---|---|---|
| Support | Reading an incoming ticket, classifying it, drafting a first reply, routing it to the right queue | Input varies enormously, output is reviewed before sending |
| Sales | Researching a company before a call and producing a one-page brief | The research route depends on what it finds; a person reads the output |
| Operations | Reading a supplier invoice or delivery note and extracting the fields into your system | Documents arrive in dozens of formats, and extraction is checkable |
| Finance | Flagging expense claims that look unusual for a human to review | Judgement call, low stakes if wrong, human decides |
| Marketing | Summarising customer interviews and clustering the themes | Unstructured input, no single correct answer, output is a draft |
Notice the pattern. In every case the input is messy, the agent produces something a person then uses, and being wrong occasionally is survivable. That is the shape that works.
The tasks agents handle badly follow the opposite pattern: uniform input, no human review, and a wrong answer that costs real money. Payment processing, contract execution and anything with a regulatory deadline belong in ordinary automation, where the steps are written down and auditable.
Where do AI agents work, and where do they fail?
Three tests decide whether AI agents for business fit a given task. A task should pass all three before you build.
1. Is the input genuinely variable? If every case looks the same, write the steps down and automate them. An agent's only advantage is handling variety, and if there is no variety you are paying for capability you cannot use.
2. Can a person check the output before it counts? Not necessarily every time, but for the cases that matter. If the answer is no, you need much stronger controls and a much narrower scope than most teams plan for.
3. Is being wrong recoverable? Ask what happens if the agent gets it wrong twenty times before anyone notices. If the answer involves money leaving the business, a regulator, or a customer relationship ending, the task is a poor first candidate however tempting it looks.
High volume plus low stakes plus reviewable output. That is the shape of every agent deployment that survives its first budget review. Teams that start with the impressive high-stakes use case are the ones providing the failure statistics.
One more filter that costs nothing: pick a task somebody currently complains about. Complaints indicate both a real cost and a person who will notice an improvement, and you will need that person at review time.
Should you build an AI agent or buy one?
Buy first, unless the process is genuinely yours. MIT's report found that purchasing from specialist vendors succeeded roughly twice as often as building internally, which is an uncomfortable finding for an agency to publish and a well-evidenced one.
| Buy a specialist tool | Build a custom agent | Hybrid | |
|---|---|---|---|
| Time to first result | Days to weeks | 2 to 4 months | Weeks |
| Upfront cost | Subscription | Build cost plus governance | Moderate |
| Fits your exact process | Partly | Yes | Mostly |
| Who maintains it | The vendor | You | Shared |
| Evidence base | Succeeded roughly twice as often in MIT's dataset | Lower success rate in the same dataset | Not separately measured |
| Your data leaves your systems | Usually yes, check the contract | No, if self-hosted | Depends |
| Best for | Common functions: support, notes, scheduling, research | Processes that are a competitive advantage, or that no vendor serves | Most mid-market companies |
| Where building is the weaker choice | This one, for any common business function. If a vendor already sells a tool for your task, building your own version means paying for development, taking on maintenance, and starting from a lower base rate of success than buying. An agency that never says this to you is not being straight |
That last cell is the honest one, and it applies to work Euracle is capable of selling. If your task is meeting notes, calendar scheduling, generic support deflection or transcription, buy the tool. Building becomes the right call when the process is something you do differently from your competitors, when your data cannot leave your infrastructure, or when nothing on the market fits.
Before commissioning any build, ask the vendor to name the off-the-shelf products they considered and why each one was rejected. An honest answer takes thirty seconds. No answer tells you something.
What does an AI agent cost to run?
AI agents for business carry three cost lines, and the one that surprises people is not the model.
The model. Agents are charged per token, roughly a chunk of a word. Cost varies per run because the agent decides how much thinking to do. A moderately complex agent running a thousand times a month typically costs a few hundred dollars in model calls. Our AI agent development cost guide breaks this down properly.
The build. A narrow agent handling one well-defined task with human review takes a few weeks. An agent with permission to act on live systems takes considerably longer, and most of that extra time is error handling and approval flows rather than the agent itself.
Governance and maintenance. This is the line that gets left out. Somebody has to review outputs, monitor cost, update the agent when the systems around it change, and investigate when it does something odd. Budget this as an ongoing staff cost, not a project cost, because it does not stop.
A useful rule when scoping: if the governance plan is shorter than the build plan, the project is not ready. The build is the part everyone enjoys writing. The governance is the part that determines whether the thing is still running in a year.
How do you measure whether an agent worked?
Count the process before you change it. Measurement is what separates AI agents for business that survive review from those that do not. This section is the direct answer to the 95% finding, and it takes about two weeks of part-time effort.
The baseline, in four numbers
Pick one process. Measure these before anything is built:
- Volume. How many times does this happen per month?
- Time. How long does one case take, end to end? Measure it, do not estimate it.
- Error rate. How often is it done wrong, and how do you find out?
- Fully loaded cost per hour of the people doing it. Salary plus employer costs plus overhead, not salary alone.
Worked example: support ticket triage
Illustrative figures. Substitute your own.
| Baseline | Value |
|---|---|
| Tickets per month | 2,000 |
| Time to triage one ticket | 4 minutes |
| Monthly triage hours | 133 |
| Fully loaded cost per hour | $35 |
| Monthly cost of triage today | $4,667 |
| Misrouted tickets | 12% |
| Average extra delay when misrouted | 6 hours |
Now the agent, after three months live:
| After | Value |
|---|---|
| Tickets triaged by agent, human-reviewed | 2,000 |
| Human review time per ticket | 1 minute |
| Monthly review hours | 33 |
| Human cost | $1,167 |
| Model and infrastructure cost | $280 |
| New monthly cost | $1,447 |
| Monthly saving | $3,220 |
| Misrouted tickets | 5% |
$3,220 a month: the difference between the two tables, which exists only because the first table exists
Against a build costing, say, $18,000 plus $700 a month in governance time, payback lands around month seven. That is a defensible number. It is also completely unavailable to any company that skipped the first table, which is the position 95% of the pilots in MIT's dataset were in at review time.
The three metrics to keep reporting
- Cost per case, which is the headline.
- Cycle time, because faster is often worth more than cheaper.
- Error rate, tracked in both directions. An agent that halves cost and doubles errors has not succeeded.
Report all three monthly, to the same person, in the same format. A project that produces a number every month is a project that does not get cancelled by surprise.
What governance do you need before launch?
Five things, all decided in writing before anything goes live. Governance is where AI agents for business are won or lost after launch. This is the section most teams skip and the one that determines whether you keep the agent.
Guardrails. Write down what the agent may do and what it may not. Which systems it can read, which it can write to, what it may never touch. Default to read-only and widen deliberately.
Human approval points. Decide which actions need a person to say yes. Anything that sends an external communication, moves money, or changes a customer record should start behind approval, even if you relax it later.
A spending cap. Agent costs vary per run, so a change in volume or a loop of retries can multiply your bill without anyone approving it. Set a hard limit that stops the agent rather than warns you.
An audit log. Record what the agent did, when, and on what basis. When something goes wrong, and it will, the difference between a twenty-minute investigation and a two-day one is whether this exists.
A named owner. One person accountable for the agent's outputs, cost and behaviour. Not a committee. Agents drift as the systems around them change, and drift without an owner is discovered by customers.
There is a sixth item that is not a control but belongs here. MIT found over 90% of organisations had employees using personal AI tools for work while only around 40% had officially licensed tools. That gap is a governance problem you already have, and it is also free research: the tasks your staff are quietly automating are the tasks worth automating properly.
How do you start? A 90-day sequence
Weeks 1 to 2: measure. Pick one process. Build the baseline table from section 7. Do not skip this and do not estimate it. If you cannot produce the four numbers, that is your finding for the fortnight and it is worth knowing.
Weeks 3 to 4: check the market. Search for an off-the-shelf tool that already does this. If one exists and fits reasonably, buy it and stop. MIT's dataset says that path succeeds roughly twice as often.
Weeks 5 to 6: scope narrowly and write the governance plan. One task, human review on everything, read-only where possible. Write the five governance items down before design begins.
Weeks 7 to 10: build the smallest version. It should do one thing, with a person checking every output. Resist adding a second task.
Weeks 11 to 12: run it alongside the existing process. Both running at once, outputs compared. This is the only way to find out whether the agent is actually better before it becomes the only option.
Week 13: report against the baseline. Cost per case, cycle time, error rate, all three against the first table. Then decide whether to widen scope, reduce human review, or stop.
Do not begin the sequence at week 7. Starting with the build is how projects arrive at week 13 with a working agent and no way to prove it. There is a step-by-step build guide in our companion piece on how to build an AI agent, but the two weeks of counting come first.
What do most teams get wrong?
No baseline. The main event, and the reason most AI agents for business get cancelled. Everything else on this list is a subset of it.
Starting with the most impressive use case. High stakes, high visibility, low tolerance for error. It is the natural instinct and it produces the failure statistics.
Scope that grows during the build. An agent asked to do one thing well becomes an agent asked to do five things adequately, usually because it was going well. Ship the one thing, measure it, then widen.
Treating governance as paperwork. The spending cap, the audit log and the approval points are what let you keep the agent after the first incident. Without them the first incident ends the programme.
Nobody named as owner. Agents degrade quietly when the systems they touch change. Without an owner, the first signal is a customer.
Measuring adoption instead of outcome. "Two hundred people used it" is not a result. Cost per case, cycle time and error rate are results.
Who pays for this: almost always the executive who approved the pilot. The cost overrun, the incident and the awkward review all land on the same desk, and the absence of a baseline means there is no defence available at that review. Agree the four baseline numbers, the spending cap and the owner before anything is built, and most of this list stops applying.
How does Euracle deploy AI agents?
Euracle approaches AI agents for business by producing a spreadsheet in the first two weeks, not a prototype. That ordering is the method, and it exists because of the 95% finding rather than in spite of it.
The Eureka Method, Euracle's discovery sprint, runs four phases.
Discover builds the baseline table: volume, cycle time, error rate and fully loaded cost for the candidate process, measured rather than estimated. It also checks the off-the-shelf market and says plainly when buying is the better call.
Design writes the governance plan, the guardrails, the approval points and the spending cap before any architecture is drawn.
Deploy ships the narrowest useful version with human review on every output, run in parallel with the existing process.
Scale reports cost per case, cycle time and error rate monthly against the original baseline, and widens scope only when those three move in the right direction together.
The stack is deliberately unexciting. Fixed sequences run on n8n or Zapier, the judgement steps call the Claude API, and the agent reads and writes through whatever is already in place, commonly HubSpot or Salesforce. Keeping the deterministic and the probabilistic layers separate is what makes the audit log meaningful, and it means the expensive component only runs where judgement is genuinely required.
Two structural commitments come from how Euracle is set up. Senior practitioners only: the people in the pitch do the work, which matters most in the two weeks of measurement where inexperience produces an estimate instead of a number. And one contract across six disciplines, so a discovery finding that the answer is a process change, an off-the-shelf subscription and no agent at all does not require a different vendor to act on. Across 50+ projects shipped for 100+ B2B teams in 12 countries, Euracle targets under 90 days from kickoff to first measurable result, and has held 98% client retention over 3 years.
If you want the baseline built before anyone quotes you an agent, that is Euracle's AI agents service. The measurement problem is sharpest for B2B SaaS companies, where support and onboarding volumes are high enough to justify an agent and rarely instrumented well enough to prove one worked.
If the conclusion is that a fixed sequence would do the job, the platform comparison in our n8n vs Zapier vs Make guide is the more useful next read.
FAQ
Conclusion
You can now approach this the way the 5% did. Pick one high-volume, low-stakes process, spend two weeks counting what it costs today in volume, time, errors and loaded rate, then check whether somebody already sells a tool that does it. If they do, buy it. If they do not, build the narrowest version with a person checking every output, run it alongside the existing process, and report cost per case, cycle time and error rate against your first table. The measurement is what separates a result from an anecdote. If you want that baseline built before anyone quotes you a build, talk to Euracle about AI agents.
Sources
- MIT NANDA initiative, The GenAI Divide: State of AI in Business 2025, published mid-2025. The 95% no-measurable-P&L finding, the $30 to $40 billion investment estimate, the learning-gap diagnosis, the build-versus-buy spread and the shadow-usage gap.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, 25 June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- n8n, official pricing page, verified 21 August 2026, for the execution-based pricing referenced in the cost section. https://n8n.io/pricing
