AI Automation and Workflows
AI in the workflow
Use language models where they are reliable, and keep them out of the places they are not. It is a defined piece of work rather than a retainer line item.
What does an AI in the workflow actually involve?
Use language models where they are reliable, and keep them out of the places they are not. In practice that means identifying the tasks where model output is genuinely good enough and verifiable, and then the harder half: building with human review at the points where errors would be costly.
Why this work exists
It is cheap to do early and expensive to retrofit, which is the opposite of how it usually gets scheduled.
AI Automation and Workflows is the wider programme; this is the piece that gets scoped first more often than any other.
The nearest piece of work to it is connecting marketing and sales data so attribution survives the handoff. Where that is already handled, this moves faster than the estimate; where it is not, the two get done together or neither of them holds.
What good looks like
You know it is holding when the numbers stop being argued about and start being used.
The gap between that and where most accounts start is why this is scoped as AI in the workflow rather than as a line inside something larger. Building the pipeline that makes reliable reporting possible is the usual second step once it is in place.
When this is the right work
What tells you to start here:
- AI tools were adopted and nobody knows whether they help
- Model output is published without review
- Costs are unpredictable and nobody owns them
If none of those is true, something else is probably the constraint, and the AI Automation and Workflows overview maps the alternatives. The usual neighbours are giving every automation a defined failure path and somebody to alert and mapping the process as performed rather than as documented.
What the work covers
Four things get done:
- Identifying the tasks where model output is genuinely good enough and verifiable. Slow, unglamorous, and the reason the rest holds up.
- Building with human review at the points where errors would be costly. Usually scoped at half the size it turns out to be.
- Prompt and output management kept in version control rather than in a document. The effort here scales with the size of the estate, not with the size of the brief.
- Cost, latency and failure handling designed in from the start, handled before the parts that depend on it rather than alongside them.
How it runs
How this gets done without the usual two-month discovery phase.
Costing the options
Where there is more than one route, each gets a rough cost and a rough timeline, so the choice is made on trade-offs rather than on whichever was described most confidently.
Removing before adding
The first pass takes things away. Adding to a setup nobody has pruned in two years makes the next person’s job harder, not easier.
Agreeing what failure looks like
The conditions under which this would be called off are written down at the start. It is uncomfortable and it prevents a year of drift.
What each part runs into
The specific resistance each part meets:
- Identifying the tasks where model output is genuinely good enough and verifiable, harder in Chicago because of a decision made by somebody who never visits the site
- Building with human review at the points where errors would be costly – in Vancouver, complicated by the fact that a procurement step nobody accounted for
- Prompt and output management kept in version control rather than in a document, which in Udaipur runs into a measurement baseline that was never taken
- Cost, latency and failure handling designed in from the start: the Noida version of this has to deal with a language mix the content has to match rather than assume
The mistake most often made here
Putting a model in a position where a confident wrong answer causes real damage, with no review step to catch it.
It is the reason this work sometimes looks like it does not work.
How it is measured
Task-level quality against the manual baseline, cost per task, and error rate. The measure is specific on purpose: a headline figure that cannot move inside the engagement is not useful.
Two people should be able to calculate it independently and get the same answer. If they cannot, it is the wrong measure.
What you end up with
What gets handed over: a documented set of AI-assisted workflows with review points, versioned prompts and cost controls.
Handed over with the working, not only the conclusion.
How it differs by sector
What this becomes in practice depends on who it is for:
- For Legal and professional services, high-value, high-intent local search under advertising rules that vary by jurisdiction. and that shapes how far this can go.
- Real estate and property brings its own limit: portals own the generic demand, so the win is locality authority and the specific listing.
- Nonprofit and NGO, constrained by the fact that two audiences with nothing in common: donors and the people served.
The industry pages carry the rest of that detail. Legal and professional services and Real estate and property are the two where this piece of work most often decides the outcome.
Common questions
Where do language models work reliably in a marketing workflow?
Classification, summarisation, extraction, translation drafting, and first drafts of structured content where a human reviews before anything is published. They work poorly as a source of fact, as a replacement for judgement, and anywhere a confident wrong answer is expensive, because the output gives no signal about its own reliability.
How do you control AI costs in an automated workflow?
Model selection by task rather than using the largest for everything, caching repeated calls, limiting context to what is needed, batching where latency allows, and hard spend caps with alerting. Costs become unpredictable when every task routes to the most capable model by default.
What comes after this?
Whatever Task-level quality against the manual baseline, cost per task, and error rate shows. The next piece is scoped from the result rather than agreed in advance, because committing to it now would be guessing. Are backlinks still important? goes through this properly, and Do AI assistants send any traffic? shows where it stops working.
Does this work at our size?
The diagnosis does, at any size. Whether the work that follows is worth the money depends on what an improvement in task-level quality against the manual baseline, cost per task, and error rate is actually worth to the business, which is the first thing to settle. A/B test explains the method. AI content policy explains why the method is the one being used.
Can this be scoped on its own?
Yes, where the diagnosis says it is the constraint. More often it runs inside a AI Automation and Workflows engagement, because the adjacent work is what makes the result hold. The scoping conversation settles which. Generative Engine Optimization for Education and training handles the common case; Google Ads Management for Education and training handles the one that catches people out.
Ask about AI in the workflow
Give us the site and one question. If the honest answer is that you do not need this, we will say so.
Find out whether this is worth doing at all
Send us your site and your goals. You will get an honest read on the gaps worth fixing first, whether or not you work with us.