The short answer: start with one workflow that happens often, takes up real staff time, runs much the same way each time, and does not depend mostly on judgement. Before you automate it, check that the process itself works. If it does not, fix it first: an automated broken process is still broken.
Four questions to rank each candidate#
List the repeated work your team does, then ask these four questions of each task. Rough answers are enough to rank them; you do not need exact numbers.
How often does it happen?
Strong candidate:Every day or every week.
Weaker candidate:A few times a year.
What does it cost you today?
Strong candidate:Skilled people spend real time on it, or it causes delays and rework.
Weaker candidate:It is quick and rarely goes wrong.
How consistent is it?
Strong candidate:Similar inputs, the same steps, a predictable output.
Weaker candidate:Every case is different.
How much judgement does it need?
Strong candidate:Little judgement: rules, templates or your own documents decide most of it.
Weaker candidate:The outcome rests on a professional's call.
The best first workflow does well on all four. A task that is frequent and costly but mostly judgement can still be helped by a draft that a person finishes; it is just not where to start automating end to end.
Rules, reading, judgement#
Many workflows mix three kinds of work, so split each one step by step:
- If it follows rules, automate it. Applying a template, moving data between systems, formatting to your house style.
- If it needs reading, AI drafts it. Summarising documents, finding the answer in your own material, writing a first draft: always from material a person has reviewed, or as a draft a person checks.
- If it needs judgement, a person decides. Giving advice, signing off, and anything that carries professional or client risk.
Automation does the assembly, not the sign-off.
Fix the process first#
If a workflow is broken (steps nobody can explain, handoffs that stall, rework caused by unclear inputs), automating it locks the problem in. That is why every engagement starts by observing how the work happens today, before choosing any model or tool: see How we work. If what I find is a process problem rather than an automation opportunity, I will say so before anything is built. Sometimes the right first step is fixing the process, cleaning up the data, or simple deterministic automation, not AI.
Workflows I have already automated#
Each is described as it would run in your firm; the “before” is the typical shape of this work, not a measured client case. In every one, a person signs off: either on each output before it is sent, or in advance, on the material or answers the system may use.
Branded reports from in-house knowledge and public data#
- Before: research is gathered, written up and formatted by hand for each report.
- After: the report is drafted from in-house knowledge plus public data, rendered in your branding as editable Office files, and a person reviews and finishes it before it goes out.
See a worked example built from public data.
Templated documents#
- Before: each document is rebuilt from the last one and reformatted by hand.
- After: each document is generated from your approved template and source content, on-brand every time, and approved by a person before it ships.
See Enterprise Document Generation, powered by DocMark.
Answers from your own documents#
- Before: staff ask a colleague or search shared drives, and a question someone already answered gets worked out again.
- After: staff or clients ask and get an answer from your reviewed material, with the source attached so it can be checked.
See Enterprise Knowledge Management.
Routine phone calls#
- Before: a routine call needs someone free to pick up.
- After: an AI receptionist answers, speaks only sentences a named person has already approved, and where no approved answer exists, says so and offers to have someone ring the caller back.
My own office line works this way: calls are answered by Citaline, the AI receptionist I am building, which is nearly ready as a product for other businesses. Until it is, I build a voice assistant for your line directly, one that answers from your reviewed knowledge with sources rather than from approved sentences: see AI Voice Assistants.
Start with one workflow#
Once you have a candidate, the AI workflow value check gives you a rough value estimate for it before a free discovery call.
Who you'll talk to
Tao Ling
Founder, Xcelerent · Sydney, Australia
Microsoft-certified data engineer · MarkLogic Certified Administrator & Developer
Value-check one workflow or request a discovery call to talk through which of your workflows to automate first.
Common questions#
What should we automate first?
Start with one workflow that happens often, takes up real staff time, runs much the same way each time, and does not depend mostly on judgement. Before you automate it, check that the process itself works.
Which tasks should stay with a person?
Anything that needs judgement: giving advice, signing off, and anything that carries professional or client risk. AI can draft the parts that need reading, but a person decides.
Should we fix a process before automating it?
Yes. An automated broken process is still broken. If Xcelerent finds a process problem rather than an automation opportunity, it says so before anything is built, and the right first step may be fixing the process or cleaning up the data.
Should we start with AI or with simple automation?
Start with what each step needs. If a step follows rules, simple automation is enough. If it needs reading, AI drafts it. If it needs judgement, a person decides. Many workflows need a mix of all three.
What has Xcelerent already automated?
Branded reports drafted from in-house knowledge and public data for a person to review, with a worked example on the site built from public data only; templated documents; answers to staff or client questions from a firm's own documents; and calls to Xcelerent's own office line, answered by an AI receptionist that uses only approved answers.