From tools to teammates

Ask what your company's software does all day, and the honest answer is that it waits: for someone to type in the deal, update the ticket, move the card. For forty years business software has been an extraordinary filing system. It remembers almost everything and does almost nothing.

The filing cabinet with a screen

Record-keeping software gives a company one version of the truth. But a salesperson has a call and then tells the CRM about it. A project manager asks five people how things are going and types their answers in so a sixth can read them.

The system records. People do. And much of what they do is feeding the system: work about work, rightly resented.

What changes when software acts

Agents reverse the arrow. The software does a first version of the work and asks a person to check it. The follow-up is drafted from the meeting notes; the ticket is sorted with a suggested reply; the overdue task is chased before anyone notices.

That is why we say teammate, carefully. A tool waits to be picked up. A teammate takes on work within a defined role and comes back when a decision is needed. The most useful agents will be specialists with a job, such as the analyst or the reviewer, because agents that fit roles are easier to trust, measure and correct than one all-knowing assistant meant to do everything.

The dynamo lesson

When electric motors became practical, many factories simply swapped the steam engine driving the central shaft for one large electric motor. The gains were disappointing. As the economic historian Paul David showed, the real benefits came only when factories were redesigned around electricity: a small motor at each machine, buildings laid out for the flow of work. It took a generation.

Most companies are at the single-large-motor stage with AI: a chat window in the CRM, a summarise button in the helpdesk. Useful, but the prize is redesigning the work so the system handles the routine flow and people handle exceptions, relationships and decisions.

The new source of power allowed a different factory, not a cheaper version of the old one. The same is true of software that acts.

The case against

Agents make mistakes: a confident, wrong email, a misread contract. And if software drafts everything, how does the junior analyst become the senior one who knows when the model is wrong?

Both objections are right about the risk and wrong about the conclusion. Companies have spent centuries building controls for fallible people: approval limits, second signatures, audit trails. Apply them to machines. An agent that can draft but not send, propose but not commit, with every action written down, is a manageable risk. As for skills, the one worth building shifts from producing a first draft to judging one, and that has to be designed into how people learn.

What a teammate needs

• A clear role. What this agent is for, and not for.
• Limits. Approval exactly where the business is committed: money, customers, prices, contracts.
• A record. Every action written where it cannot be quietly edited.
• Memory. So the service agent knows what the sales agent promised.
• A manager. Someone who reviews the work and widens or narrows the limits.

It is how a well-run company treats a new hire, made the default in software.

What this means for a mid-market leader

Judge AI by the work it removes, not the features it adds. If nobody can say which recurring tasks your people will stop doing, the AI is decoration.

Write down your decision rules first. Which actions commit money, customers or reputation, and who may approve each? Most companies have never made this explicit.

Redesign one process end to end. Pick one your people dislike feeding, such as the weekly status, let agents take it on under clear rules, measure the hours returned, then move to the next.

Business software spent forty years learning to remember. The next decade is about teaching it to act, carefully, under people who keep the decisions.