Which Agency Tasks AI Handles Well, and Which It Doesn't

AI handles agency tasks well when the source material is clear, the output is easy to check and a mistake can be caught before publication. Start with brief summaries, content variations and report drafts. Keep positioning, budget decisions, sensitive client conversations and final approval with an accountable person.

The useful unit is a task, not a job title. An account manager may delegate ten minutes of meeting-note sorting while keeping the difficult conversation about a missed target. A copywriter may request headline options while retaining the argument, evidence and final wording. Judge the complete handover, including review time.

Follow me on Instagram@sagnikteaches

Separate preparation, recommendations and authority

For a marketing agency, three different things often get called automation. Preparation turns existing information into a useful form. Recommendation proposes what to do. Authority means the system can spend money, publish, contact somebody or change a live account. These need different levels of trust.

Connect on LinkedInSagnik Bhattacharya

A report draft belongs in preparation. Suggesting that the client stop a campaign is a recommendation. Pausing that campaign is an exercise of authority. A convincing paragraph does not prove that the system is ready for the other two jobs.

Subscribe on YouTube@codingliquids

Use the table below as a starting allocation, then adjust it to your client contracts and team experience. These are suggested working boundaries, not claims that a particular product can perform every task.

Before putting a task in the first column of your own list, ask who can check it. If your junior account executive must ask a senior strategist to rework every output, the time has moved between desks. It has not necessarily disappeared. Include that senior person's time in your trial.

Where clear inputs make AI useful

Briefs: extract gaps before drafting anything

Ask AI to separate confirmed requirements from suggestions and missing decisions. This works best when you provide a short, agreed source pack: the enquiry, the latest meeting notes and the signed scope. Tell it which source wins if they disagree.

Consider an illustrative agency brief for a packaging supplier: “Promote the new boxes. Sales wants more enquiries. Make the offer stronger. We might include free delivery.” A useful output identifies the product range, the intended buyer and the unapproved delivery offer. It does not turn “might” into a headline promise.

The account manager's next message could be: “Please confirm the minimum order, approved delivery wording and whether the enquiry form is for samples or quotations.” That is a valuable result even though no copy has been written. It prevents a polished draft from hiding an unfinished commercial decision.

Variants: change the expression, hold the facts still

AI is a useful candidate for producing several openings, email subjects or short descriptions from one approved message. Give it exact product facts, a length limit and examples of what the client dislikes. Keep the source text beside the output during review.

For an illustrative spare-parts manufacturer, the approved fact is “Orders placed by noon are dispatched within two working days.” A draft that says “Get replacement parts in two days” has changed dispatch into delivery. Reject it even if it reads more smoothly. The distinction belongs in the reusable instructions.

Produce four opening lines for a spare-parts enquiry email.
Approved fact: orders placed by noon are dispatched within two working days.
Audience: maintenance managers comparing replacement components.
Do not promise delivery dates, stock availability or compatibility.
Keep each option under 18 words.
After each option, identify the approved fact it uses.

An illustrative output might be: “Need a replacement? Order by noon for dispatch within two working days.” The fact is preserved, but “a replacement” is vague. The writer could improve it to “Ordering a replacement drive belt? Order by noon for dispatch within two working days,” provided the approved offer covers that product. Choosing the relevant detail is still editorial work.

Once this narrow step works, the agency content workflow from brief to approval helps place it between briefing, drafting and client sign-off. Do not let a good headline trial quietly turn into automatic publication.

Research: make a reading list easier to inspect

Use AI to compare documents you can open and check. Ask it to list the source for each observation, separate quotations from paraphrases and say where evidence is missing. Keep facts from the client's own records distinct from statements made by competitors.

For example, an agency reviewing six publicly available packaging catalogues could ask for a table of stated minimum quantities and ordering steps. A missing quantity should appear as “not stated in supplied source”. It should not become an estimated minimum based on what similar suppliers tend to offer.

Keep interviews, source selection and conclusions human. Six catalogues can show how six businesses describe themselves. They cannot establish what every buyer wants. A useful research assistant saves searching within documents; it does not turn a small source pack into market-wide evidence.

Reporting: describe the numbers before explaining them

Give the assistant a checked table, agreed metric definitions and the reporting period. Ask first for observations. Request possible explanations in a separate section, each accompanied by the evidence needed to test it. Calculate totals and percentages in your spreadsheet before drafting the narrative.

An illustrative report says enquiries rose from 40 to 52. That is an increase of 12, or 30%. If spend also rose, “our new creative improved efficiency” is not established by the enquiry count. The report needs spend and lead-quality information before making that claim.

The difference is practical. “Enquiries increased by 30%; we are checking whether the extra enquiries meet the sales team's criteria” is useful. “The campaign is now 30% more effective” smuggles in a judgement the figures do not support.

Administration: draft the reminder, verify the commitment

Meeting summaries, production checklists and approval chasers are good trial candidates when the record is clear. Ask for actions with an owner, a due date and the sentence that supports each entry. Use “owner not agreed” or “date not agreed” rather than allowing the system to fill gaps.

In an illustrative client call, “We could review the landing page on Thursday” becomes “Client approval due Thursday” in the AI summary. That changes a possibility into a commitment. The account manager should correct it to “Ask whether Thursday works for a review” before sending the notes.

Keep repeat reminders tied to the live approval record. A draft may be accurate when created and wrong a day later if the client has already replied. Before sending, check the recipient, latest thread and current status. Faster chasers are only helpful when they chase the right thing.

Six hours recovered from monthly reporting: an illustration

A six-person marketing agency produces 12 client reports each month. Its current process takes 65 minutes per report: assembling the figures, writing the commentary and reviewing the result. The agency selects commentary drafting as its first AI trial because the numbers already come from a checked spreadsheet.

It starts with three old reports whose final versions are available. The account lead removes client identifiers and commercial details that are unnecessary for the test. A writer prepares a short reporting brief: explain changes plainly, distinguish observation from possible cause, and never invent a sales outcome.

For the live pilot, each report still needs eight minutes of preparation, twelve minutes of drafting and correction, and ten minutes of senior review. Those are illustrative measured times for this example, not an industry benchmark. Another agency should time its own work.

The agency gives the two people running the pilot Claude Team Standard seats. The USD list price is $25 per seat a month on monthly billing, so two seats cost $50 a month. The plan has a two-seat minimum. Content on Claude Team is not used for model training by default; the agency still needs an approved basis for putting client material into the service.

At an illustrative internal time value of $35 an hour, six hours represents $210 of capacity. Subtracting $50 leaves $160 of monthly value before other overheads. Four hours of initial setup adds $140 of staff time in the first month, leaving only $20 on that simple first-month comparison.

Those figures describe capacity, not cash automatically added to the bank. The agency benefits if the recovered time reduces overtime, avoids extra freelance work or supports work it can actually sell. If it simply produces longer reports that clients do not read, the trial may have improved neither margin nor service.

Quality changes the result too. Suppose two reports include unsupported explanations for a drop in enquiries. The lead removes them and adds a required “evidence still needed” field. If the same error keeps returning, they restrict AI to observations and write the interpretation themselves.

They retain the original sheet, draft, corrections and approved report for each test. At the end of the month, they compare total time, corrections per report and client questions caused by unclear wording. The client-reporting tutorial develops this specific workflow further.

Keep these agency decisions with people

Positioning requires a choice somebody can defend

AI can propose three positioning options from a supplied brief. Your strategist should decide which audience to prioritise and what the client is willing to stop saying. A list of plausible alternatives is not a decision about the business.

For an illustrative laboratory client, “fastest results” may sound attractive. Yet the operations team may value careful handling of unusual samples more than speed. The agency needs that conversation before choosing the promise. Ask AI for questions to test each position, then put those questions to the people who can answer them.

Creative judgement includes what to leave out

A tool can generate a large number of campaign ideas. The agency still needs someone to decide which one fits the brief, avoids confusing the audience and can be delivered within the budget. Do not ask a client to perform that selection by sending every generated option.

An illustrative home-care provider wants an enquiry campaign that feels reassuring. A generated concept based on a family emergency may attract attention but be wrong for the client's preferred tone. The creative lead should reject the premise, not merely soften the wording. Editing individual sentences cannot repair an unsuitable idea.

Client conflict needs context and accountability

AI can help prepare a calm draft when a client questions an invoice or rejects work. It should not decide what the agency admits, refunds or promises. Those choices depend on the agreement, the relationship and facts that may never appear in the email thread.

For example, “We accept that our work failed to meet the brief” is materially different from “I understand why the draft was disappointing.” An account director should decide which statement is accurate. Do not let a request for a warmer tone become an unreviewed concession.

Budgets, claims and final release need named approvers

Keep campaign spending limits and publication permissions outside the assistant's free-form suggestions. A proposed move from one channel to another should identify the amount, reason, evidence and approving person. A draft should stay a draft until that person accepts it.

The same applies to product claims. If a packaging client has approved wording about recycled material, the writer must check the exact scope before expanding it into a wider environmental claim. Where claims raise legal or technical questions, ask the relevant qualified adviser. Client enthusiasm is not evidence.

Use a named reviewer for facts and another, where needed, for release. The person checking a delivery promise may be the client's operations manager, while the agency account lead controls publication. Making these responsibilities visible prevents “I thought you had checked it”.

Choose a pilot with three gates

You do not need a large scoring exercise to select the first task. Spend 30 minutes with the people doing the work and apply these gates to ten recent examples. Treat the numbers below as proposed trial rules that you can tighten for your clients.

1. Evidence gate: can the reviewer find the supporting source for every important claim? If no, improve the source pack before testing generation.
2. Review gate: can a competent person check the output in less time than producing it manually? Record the time instead of guessing.
3. Consequence gate: can an error be stopped before it reaches the client, public or spending account? If no, narrow the task to preparation.

Here is an illustrative filled-in pilot record: “Task: monthly report commentary. Owner: account lead. Inputs: approved figures and metric definitions. Output: draft observations and questions. Forbidden: causes stated without evidence, forecasts and spending changes. Reviewer: strategist. Trial: ten reports. Stop condition: any confidential material from another client appears.”

A suggested release target is that all ten reports preserve every checked figure and contain no unsupported claims after review, while total handling time falls. Track drafts needing major rewriting separately. A result that only becomes acceptable after a full rewrite should not be counted as an efficient success.

Agree what happens when the assistant cannot complete the task. The fallback might be the existing report template and a manual note to the account lead. Missing source data should create an exception, not an empty but reassuring paragraph. For a wider measurement method, see checking AI quality as well as speed.

Make the boundary visible in everyday production

Give each recurring AI task a small instruction sheet. Include the permitted input, expected output, source of truth, reviewer and actions the tool is not allowed to take. Store the approved example beside it. A new colleague should be able to understand the handover without reading an entire chat history.

Separate clients in your working folders and approved AI work areas. Do not put one client's prices, strategy or internal documents into another client's prompt as an example. Replace reusable patterns with fictional details, and check that distinctive commercial information has also been removed.

Keep one short correction log. “Dispatch became delivery” is a better entry than “AI got the tone wrong”. It tells the next reviewer what to look for and gives the owner a precise instruction to improve. If the same correction appears repeatedly, reconsider whether the task belongs in the pilot.

Measure waiting time separately from working time. If a draft takes 20 minutes instead of an hour but sits with the client for nine days, the delivery date has barely changed. In an illustrative agency, two named client approvers disagree about an offer and neither owns the final answer. Producing more variants adds work. The useful intervention is to agree one approver and a deadline before the next draft. AI may speed production; it cannot settle that responsibility for you.

Finally, review the handover when the client changes the offer, audience or approval rules. A previously sound instruction can become stale. The goal is a repeatable piece of agency work with a clear owner, a checkable result and enough recovered time to justify keeping it.

Further reads

• AI Rollout Plan for a Ten-Person Marketing Agency — Turn a successful task trial into a team rollout.
• Hire an AI Specialist or Upskill Your Agency Team? — Decide who should own the new workflow.
• How Small Agencies Use AI to Write New-Business Pitches — Apply the boundaries to proposals and pitches.
• AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — Agree disclosure and approval responsibilities with clients.
• How Agencies Keep Brand Voice Consistent in AI-Written Content — Keep each client's approved voice separate.
• Which Marketing Tasks Should a Small Business Never Hand to AI? — Fourteen marketing tasks where AI may draft but must never decide or publish, each with a real-looking failure, plus a sign-off policy you can copy.
• How to Turn Meeting Notes Into Tasks Automatically With AI — A capture, extract, approve and push routine that turns every call's action items into assigned, dated tasks without anyone typing them up.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.