What does an AI consultant do?

Most people hear “AI consultant” and think the job is recommending tools.

That is part of it sometimes, but it is not the main job.

The real job is understanding what the business is trying to improve, deciding whether AI belongs in the solution, and making sure whatever gets built can work inside the actual operation.

Sometimes that requires AI. Sometimes it requires normal automation, systems integration, a process change, or simply configuring software the company already owns.

What does an AI consultant actually do?

1. Define the business problem

Before discussing models, the consultant should understand the work:

• What starts the process?
• Who owns it?
• Which systems and information are involved?
• Where does work wait, repeat, or break?
• What exceptions require human judgment?
• What is the problem costing the business?

The output should be a defined problem, a process boundary, an accountable owner, and a baseline that can be measured later.

2. Assess the process, data, and systems

AI depends on context. In a company, that context may be spread across databases, documents, email, CRM records, support tickets, call notes, project tools, and the knowledge employees carry in their heads.

The consultant needs to identify which sources matter, whether they are usable, who owns them, who is allowed to access them, and how they connect to the workflow.

On one Losung project, a contracting company came to us about improving how quotes were sent to customers. While examining the process, we found useful sales and customer information scattered across call notes, support conversations, and internal discussions.

Once that information was cleaned and organized, it became an internal knowledge resource and exposed a pattern between delayed quote follow-up and lost work.

The original request was about quotes. The wider problem was that the company could not use the information it was already producing.

That is why discovery matters.

3. Decide whether AI is the right mechanism

AI can help when the work involves language, documents, images, variable inputs, classification, extraction, search, summarization, or drafting.

Traditional automation is normally better when the rule can be written clearly: validate these fields, create this record, route it to this owner, and escalate if nothing happens within two days.

Many useful systems combine both. AI interprets the messy input; deterministic rules control what the business does next.

A responsible recommendation can also be “not yet.” If the process changes every week, the information is unreliable, or nobody owns the result, AI will not repair that.

4. Design, implement, and test the system

The model is only one part of an implementation. The complete system may also need integrations, permissions, validation, human review, exception handling, monitoring, and a usable interface.

Testing should include the difficult cases: missing information, conflicting records, unusual inputs, failed connections, and uncertain outputs.

Anybody can make the happy path look good in a demonstration. Production is where the awkward cases show up.

5. Establish controls and ownership

The consultant should define what the AI may do, what a person must review, which information it may access, what happens when confidence is low, and who approves changes.

The team also needs training, documentation, and a named owner after launch. A system is not implemented merely because the technology works.

6. Measure whether the business improved

The right measure depends on the process. It might be time per case, backlog, error rate, cost, throughput, conversion, employee capacity, or customer experience.

The baseline must be recorded before the process changes. Otherwise, everybody is left saying the system “feels faster,” which is not the same as proving value.

What should an AI consulting engagement leave behind?

Output	What it should answer
Defined business problem	What are we improving, for whom, and why?
Current-state process map	How does the work happen now, including exceptions?
Data and systems map	Which sources, permissions, integrations, and owners are involved?
Options assessment	Why is AI better than a simpler alternative?
Working system or executable roadmap	What will actually exist at the end?
Measurement and ownership plan	How will results be measured, and who runs it afterward?

If a proposal cannot answer those questions, it is probably too early to approve the build.

At Losung Consult, we work across AI, automation, systems integration, data, and custom software because the business problem normally crosses those boundaries.

Start with the work that is stuck. We can work out which technology—if any—belongs in the answer.