How to leverage consultants in the age of AI

AI gives consultants a much better starting point.

Before the first workshop, a consultant can already have a market scan, a first pass through the data, a few hypotheses and possibly a rough prototype. After ten interviews, the notes can be compared that afternoon while the details are still fresh.

As a consultant and an AI power user, I see that as good news. It leaves more time for the work that actually justifies bringing someone in: deciding what the evidence means, challenging the comfortable answer and putting a recommendation on the table.

In this article, I want to separate two aspects of AI that often get reported under the same heading:

AI for alpha, where AI helps you make a better bet.

AI for automation, where AI helps you run the current work with less effort and cost.

Saving 10,000 hours and making one better investment decision are both valuable. They are not the same kind of value.

1. AI for alpha

Ask a model to develop a growth strategy for a company using its annual report and other public information.

The answer will probably be decent.

It may suggest entering a new market, changing the pricing model, improving retention, adding a product or acquiring a smaller player.

The problem is not that these ideas are wrong. The problem is that another company can run the same exercise before lunch.

A broad question, public information and a widely available model will usually produce a broadly sensible answer. There is nowhere for a company-specific insight to come from.

That insight has to enter through the people who know the company.

An employee notices that customers who ask for a particular integration tend to leave six months later.

A model can test that observation against support tickets, implementation data and churn cohorts.

A consultant might then speak to the sales team and discover that the integration itself is not the problem. It is a sign that the company is selling to customers the product was never designed to serve.

The useful idea did not magically appear because the model read more documents.

It started with someone close enough to the work to notice that something felt off.

The same thing happens with internal processes.

A model sees that an approval takes nine days and suggests automating it.

The people doing the work know that the approval is slow because nobody trusts the data entering the process. Some cases are checked three times because the source system is often wrong.

Automating the approval would make the process faster. It would also make the mistakes faster.

This is why I would not begin an important strategy exercise by asking an agent:

What should this company do?

Start with the judgment already inside the company.

What does the sales team suspect?

What does the operations team keep working around?

Which customer behaviour has never made sense?

Where does management believe money is being lost?

A consultant can turn those observations into hypotheses. The agent can then search the data, compare explanations and try to disprove them.

Do not wait for the agent to decide what your company should care about.

It has no reason to care about the strange movement in one metric, the exception that appears every Friday or the difference between the official process and what people actually do. Someone has to point it in that direction.

Faster, better and sometimes more wrong

A 2026 Organization Science study followed 758 management consultants working on realistic assignments.

On 18 tasks that fell within the tested model’s capabilities, consultants using AI completed 12.2% more tasks and worked 25.1% faster. Their output was also assessed as higher quality.

The researchers then gave participants one complex managerial problem deliberately chosen to sit outside the model’s capabilities. Consultants using AI were 19 percentage points less likely to reach the correct answer.

Their recommendations still sounded better. Even when the underlying answer was wrong, the AI-assisted work was rated as more coherent and persuasive.

That is a useful warning.

AI improves the answer when it is capable of doing the work. It can also improve the packaging when it is not.

A weak recommendation no longer has to look weak.

It can arrive with a clean storyline, supporting arguments and the confidence of someone who has never had to defend the decision after it failed.

This is where a consultant should earn the fee.

The review cannot consist of checking whether the slides are clear and the calculations add up. Someone has to ask which assumption is carrying the recommendation, which evidence contradicts it and what information never reached the model.

A serious recommendation should also include the fact most likely to kill it.

If the answer depends on customer retention remaining above 85%, say so. If the investment case collapses when implementation takes nine months instead of six, show it. If the strategy only works because three departments are expected to cooperate in a way they never have before, that belongs on the first page.

AI can give a consultant more time to investigate those questions.

It should not give the consultant an excuse to stop asking them.

Your people already hold part of your alpha

Companies tend to think of their proprietary advantage as data, technology, intellectual property or distribution.

A large part of it also sits inside people’s heads.

The account manager who knows which customer complaint signals a real risk.

The operations lead who can tell the difference between an unusual case and the beginning of a larger problem.

The employee who knows that a KPI looks good because one team quietly fixes the underlying data every month.

The consultant who has seen a similar plan fail elsewhere and recognises the same assumption hiding inside the new version.

None of this is useful at scale if it remains an anecdote shared during a meeting and forgotten two weeks later.

The judgment needs to enter the system.

When a person overrides an agent, record why.

When the same exception appears repeatedly, question the rule instead of teaching the agent to handle an endless list of exceptions.

When a consultant finds that a metric is misleading, update the context used by future agents. Otherwise the next agent will confidently repeat the same mistake.

The model is unlikely to be proprietary.

Your judgment can be.

2. AI for automation

Automation is easier to measure.

Suppose a monthly report takes 80 hours to produce. An agent reduces that to eight.

That is a clean business case. Do it.

A person copying information between two systems every morning is not holding the company together through irreplaceable human creativity. The work can go.

Still, optimization should come after deciding what deserves to exist.

If nobody reads the monthly report, generating it in eight hours is not much of an achievement.

Generating it in 30 seconds is not an achievement either.

You have made the useless report arrive earlier.

Every automation exercise should be allowed to end with: Stop doing it.

There is also a difference between reducing cost and creating alpha.

If several companies automate the same back-office process and each reduces its cost by 20%, they all benefit.

Over time, some of that saving will probably appear in lower prices, faster service or more investment elsewhere. The lower cost becomes the new baseline for the market.

That is good management.

It does not automatically make one company harder to compete with.

And if you follow the productivity argument far enough, the organization starts to look slightly strange.

Twenty people become ten. Ten become three people supervising 30 agents. Eventually, agents are opening support tickets for other agents while one human approves permissions.

If there are no employees left, is it even a company anymore?

Should companies rebuild their software?

This is where I have more of a nuanced answer.

Companies pay a lot for software and often use a small part of it.

Take a product costing €30 per employee per month across 1,000 employees:

€30 × 1,000 × 12 = €360,000 per year.

Over five years, that is €1.8 million before any price increase.

Maybe employees use four screens.

At that price, asking why you cannot build those four screens internally is no longer a ridiculous question.

Coding agents have made the first version much cheaper. A company can build a smaller tool that follows its actual process, connects directly to its systems and avoids paying for dozens of unused features.

The calculation looks obvious until the tool goes live.

Then the authentication system changes.

One country needs a different workflow.

The audit team asks for a history that was never designed.

The employee who requested the tool moves to another role.

Finance wants one approval rule. Operations needs an exception. Security wants the exception removed.

A coding agent can update the integration and fix the tests.

Someone still has to decide what the software should do.

That is the part people skip when they compare a subscription invoice with the cost of generating version one.

The real comparison is closer to:

Avoided subscription cost − build − maintenance − security − support − change − eventual retirement

The subscription may also be buying something that is easy to overlook: the right to call someone else when the product breaks.

Before replacing a subscription, name the person who will own the internal tool for the next five years.

“An agent will maintain it” does not answer the ownership question.

The agent can keep the software aligned with its instructions.

Who keeps the instructions aligned with the company?

SaaS may lose the interface

My current view is that SaaS will not simply disappear, the core is still important.

Some of it will become less visible.

Most business software was designed around a human opening an application, moving through several screens and clicking a button.

An agent does not necessarily need the screens.

It needs the underlying capability:

Create the invoice.

Retrieve the contract.

Update the forecast.

Open the support case.

Schedule the shipment.

MCP, or Model Context Protocol, provides a standard way for AI applications to discover and use tools, retrieve contextual data and access reusable workflows. An MCP server can expose actions and information that an agent can call without navigating the product like a human user.

I suspect a growing part of SaaS will move in that direction.

The software company still operates the reliable service underneath. It handles uptime, security, regulatory changes and difficult edge cases.

The agent accesses those capabilities through MCP or an API and uses them inside a workflow specific to the company.

That gives companies a third option.

They do not have to buy an enormous product and force every employee through its interface.

They also do not have to rebuild the entire product internally.

They can buy the commodity capability and own the small layer that contains their judgment.

There is little advantage in building another login page, date picker or permissions system.

The interesting part is deciding:

Which supplier requires additional review?

Which customer qualifies for an exception?

Which source should be trusted when two systems disagree?

When should the workflow stop and involve a person?

Those rules may be specific to the company. They reflect its risk appetite, customer promises and operating experience.

That is the layer worth owning.

Buy the capability.

Own the judgment.

Agents are workers, not coworkers

I would also stop calling agents “coworkers.”

A coworker is assumed to have broad context, personal initiative and some discretion over what deserves attention.

Most company agents should look more like workers with defined jobs.

For example:

Review new supplier invoices against purchase orders and signed contracts. Use only the finance system and contract repository. Flag mismatches above €5,000. Do not approve payments. Send unresolved cases to the financial controller.

That agent has a job.

It knows what information it may use, which actions it may take, where its authority ends and who receives the exception.

Compare that with:

Review our finance operations and find opportunities.

The second instruction may produce an impressive memo. It may also produce a collection of ideas that could apply to almost any finance department.

Companies should not build their operating model around waiting for agents to have ideas.

Agents can spot anomalies and make suggestions while doing their work. That is useful.

Direction still needs an owner.

An employee or consultant forms a view based on something observed in the business. The agent investigates it. A person decides whether the evidence is strong enough to act.

The agent can do a large amount of work between the first question and the final decision.

It should still know where its job ends.

What to expect from consultants as of today

Research took time. Models took time. Comparing interviews took time. Turning everything into a clear document took time.

AI compresses much of that work.

Companies should use the capacity it releases, give consultants a decision, not a topic.

“Analyse this market” will produce a market analysis.

“Should we enter this market before 2028, and what would have to be true for the investment to work?” forces a recommendation.

Give them access to the employees closest to the work, including the people who disagree with the official explanation.

Ask them which hypotheses they rejected and which fact changed their view.

Use them to decide which processes should disappear, which software capabilities should be bought and which decision rules are specific enough to own.

Then capture the reasoning.

A useful consulting output in the age of agents is more than the final recommendation. It includes the evidence behind it, the assumptions that could break it, the exceptions that require human review and the point at which the decision should be reconsidered.

Employees bring the daily reality of the company.

Consultants bring an outside view and the willingness to question things that have become normal internally.

That combined judgment is part of the company’s alpha.

Every company will have capable models. Many will connect their agents to the same software capabilities.

The difference will sit in the questions they ask, the context they trust, the actions they permit and the exceptions they refuse to automate.

Protect your company’s alpha by using the judgment of your employees and consultants.

Let agents carry that judgment through the work and do not wait for them to invent it.

When your agents act, whose judgment are they extending?