Productivity assistants or custom AI systems: what is each one for?

The licenses arrive: everyone now has a general AI assistant in their email, documents and chat. Within a month, a department head asks whether that means the returns project can be stopped.

It can’t, because the two do different jobs. General assistants make individuals faster at drafting, summarizing and searching. Custom systems change a workflow, with access to its data, approval points and measurement against a baseline. You will probably need both, bought and built for different jobs.

What do general AI assistants do well?

General assistants, such as Microsoft 365 Copilot, ChatGPT Enterprise or Gemini, are built for the individual. They are good at:

• Drafting and rewriting: a first version of an email, a proposal section, a summary of a policy.
• Summarizing: a long thread, a meeting, a report nobody has time to read.
• Finding things: searching the files and messages a person can already open.
• Thinking aloud: outlining, brainstorming, testing an argument.

The person brings the context, judges the output and carries it into the next system. That is the design: the assistant is broad, and the person is both the integration and the reviewer.

Where do they stop being enough?

When the work is a workflow rather than a task: repeated cases, rules to apply, several systems to consult, decisions that commit money or reach customers. The warning signs are easy to spot:

• people paste data from three systems into a chat window to get one answer;
• everyone has their own prompt for the same task, and results differ by person;
• sensitive data travels into prompts outside any approved flow;
• nobody can say whether errors went down, because nothing is measured against a baseline.

At that point the assistant is being asked to be a system, without the data access, controls or measurement a system needs.

What does a custom system add?

A custom system is built around one job. It reads the workflow’s own systems with mapped permissions, applies your rules in a fixed sequence, writes drafts where the work happens, stops at approval points and records what it did. Its quality is tested on an evaluation set before launch and monitored after. That is what custom AI software means in practice.

An assistant makes a person faster. A custom system changes how the work runs.

How do you decide for a given task?

Four questions sort most tasks:

1. Is it one person’s task, or a queue of similar cases?
2. Does it need data from systems the person would otherwise copy by hand?
3. Does the output commit money, reach customers or need to be the same whoever handles it?
4. Is the volume high enough to repay building, testing and monitoring?

If the answers point to one person, no system data, nothing committed and low volume, use the assistant. If they point to a queue, several systems, consequential output and real volume, design a workflow system, bought if a product fits and built if not, as build, buy or wait explains. Some tasks sit in between, such as a weekly report drafted from three systems. Start with the assistant and a shared prompt; if the task keeps growing, it has become a system.

Take an illustrative contrast. Drafting one email to a supplier about a late delivery is an assistant’s job: the person reads, edits and sends, and saves a few minutes. Preparing decisions on hundreds of returns a day is not. Each case needs the order record, the photos and the policy paragraph that applies, the same answer in every store and channel, an approval step and a record of who decided. That is a system, like the illustrative returns engagement.

The two are measured differently, too. An assistant’s value is individual time saved, spread thinly across many people; sample it on specific tasks instead of asking people to estimate it, and don’t expect it to show up as a budget line. A custom system is measured on the workflow: time per case, errors, cost per case and decision time, against the baseline recorded before launch.

Can the two work together?

Yes, and they should. The assistant is the front door for individuals, and the custom system does the workflow behind it; through a connector, someone in customer service could ask the assistant why a return was refused and get the system’s own record and citation. Both can share the same foundations: identity and permissions, approved model providers and one acceptable use policy.

The assistant also feeds the pipeline. When a team reuses the same clever prompt every day for the same task, that task is a candidate for a proper workflow project, which is the healthy end of shadow AI. People who use an assistant daily also make better reviewers for custom systems: they already know where a model is fluent and where it needs checking.

What should IT set up either way?

• Data settings. Retention, logging and whether prompts may be used to train anything under the provider’s enterprise terms. Take those questions to your data protection officer.
• Access. Who gets which tool, through single sign-on, and what an assistant may search. One that searches everything a person can open will surface every file shared too widely, so fix permissions before switching search on.
• An acceptable use policy. Which tools, which data, who checks the output. A short policy people will read beats a long one they won’t.
• Training. Role-based, on real tasks, with the limits made clear. Licenses without training and new routines tend to stall, which is the problem AI enablement addresses.
• An owner per tool. Someone owns its settings, its policy and the renewal decision, made on measured use rather than on the number of licenses.

Two lists, two answers

List two things: the tasks people bring to the assistant most often, and the queues where similar cases pile up. The first list is for training and shared prompts. The second is for AI strategy and discovery, which sorts candidates into build, buy, wait or stop.