Our last article argued that the biggest AI gains come from redesigning work, not speeding it up. This is what that redesign looks like on a real process - and why it is now something you can build, not just describe.

The shape of most processes today

Pick almost any end-to-end process in a business and it has the same shape: a chain. A request comes in, someone does the first step and passes it on, the next person picks it up when they get to it, does their part, and passes it on again. Each step waits for the one before it to finish.

The striking thing, when you measure these processes, is how little of the elapsed time is real work. A task that takes a few hours of effort can take days to complete, because most of the calendar is spent sitting in queues: waiting for a colleague to reach their inbox, waiting for an approval, waiting for someone to confirm a number that another system already holds. The work is fast. The waiting is slow. And the waiting exists because each step needs a person to start it.

A worked example, before

Consider a large seasonal order from a major retail customer, in a business that sells several brands. Today it might run like this:

• A commercial coordinator receives the order and checks it against the account’s terms.

• Once that is done, planning checks whether the quantities are available across the relevant brands - often across more than one system.

• If something is short, the order pauses while options are worked out: substitute, split the shipment, or pull from another region.

• Logistics then schedules fulfilment and confirms dates.

• Finance sets up billing and runs credit checks.

• Customer service confirms back to the retailer.

Each of those steps is competent. The problem is that they happen one after another, and between each one the order sits in a queue. A shortfall discovered at the availability step sends the whole thing backwards. The retailer hears nothing until the chain reaches the end. A process that involves a few hours of real decision-making can take the best part of a week.

The same process, reimagined

Now redesign it for people and agents working together, and the chain breaks apart into parallel workstreams.

The moment the order arrives, several things start at once. An availability check runs across every relevant brand and system simultaneously, rather than waiting its turn. A credit and terms check runs alongside it. A logistics feasibility check begins on the assumption the order will proceed, so dates are ready the moment it is confirmed. None of these has to wait for the others, because none of them needs a person to press start.

Where the steps were pure handoff - moving information from one person or system to the next, re-keying it, confirming it - they simply disappear. The information is already shared.

When something needs a decision rather than a lookup - a real shortfall, an unusual credit position, a customer-specific exception - that is where a person comes in. Instead of working the whole chain, the commercial lead is handed a clear, assembled choice: here is the shortfall, here are three viable options with their cost and date implications, which do you want? The judgement that used to be buried under hours of coordination becomes the main thing the person does.

It is not a faster version of the old chain. It is a different shape.

What changed, and why it works

Four shifts are doing the work. Steps that ran in sequence now run in parallel. Steps that existed only to move information are eliminated rather than automated. People move from running the process to deciding the things that need judgement. And the connective tissue between steps - the chasing, the assembling, the confirming - is carried by software. This is the move from a strictly sequential process to overlapping workstreams.

Why this is buildable now, not just describable What changed, and why it works

For years this was a whiteboard exercise, because the software could not carry a process across systems on its own. That has changed. The enterprise platforms most organisations already run now support multiple agents that each handle a specialist part of a process and coordinate with one another, including across different systems. SAP’s Joule platform, for example, supports agents that collaborate on a multi-step process and an agent-to-agent protocol for them to communicate; Google Cloud’s Gemini Enterprise Agent Platform provides a managed way to build and orchestrate multiple agents. Both support the open protocols, including Model Context Protocol, that let agents read from and act on existing systems such as the ERP without replacing it. Specific capabilities sit at different stages of release, so the exact feature set is worth confirming at the time of reading.

A common and effective pattern is an orchestrator agent that coordinates several specialist agents - one for availability, one for credit, one for logistics - while the people involved supervise and make the calls that matter. That is the parallel workflow above, made real.

Where people fit

It is worth being clear about this, because the instinct is to assume the people disappear. They do not. What changes is what they spend their time on. The hours that went into chasing, re-keying and waiting are gone; the time goes instead into the decisions the process used to bury. Done well the role becomes more skilled, not less: people supervise a small team of agents, handle the exceptions, and own the outcome. That also means governance belongs in the design from the start - agents need clear boundaries, oversight and escalation, the same as any team.

Start with one process

You do not redesign the whole business at once. You take a single process that is visibly slow because of its queues, map it end to end, and ask which steps are decisions and which are just handoffs. The handoffs are your parallel workstreams and your candidates for removal; the decisions are where your people stay. Redesign that one process, build the agents to support it on the systems you already run, and you have a working example that makes the next one far easier to picture.

In our last article we argued that the biggest gains come from redesigning the work rather than speeding it up. This is what that redesign looks like in practice. If you would like to see it mapped onto one of your own processes, that is exactly the kind of session we would start with. To know more, contact us.