Your next employee apparently needs no desk, never takes leave, and is delighted to work through the weekend.
Before you put an AI agent on the Christmas party list, there is a small matter of its job description.
Picture a hotel manager in Westlands opening Monday's inbox. There are enquiries about conference packages, a half-finished campaign report, and someone asking whether the room rate includes breakfast. The manager doesn't need a digital colleague with an impressive introduction. She needs those jobs finished, without spending the afternoon supervising the help.
That’s the allure of AI agents. They promise to move work along, rather than give you another answer to copy and paste. But is the commercial value real, or have we reached the point where your business needs one because the business next door says it has seven?
There are real results worth examining. Klarna estimates it saved $10 million from AI-powered marketing in 2024. A digital agency reports 32% lower operational costs after introducing an analytics system with agents. Salesforce reports lead-nurturing agents contributed to 1,500 closed deals.
Those results describe different jobs and different kinds of AI. Understanding the difference is how you avoid buying the story and missing the value.
First, what job are you actually hiring for?
An automation follows a process you have set. An AI agent can choose what to do next within the tools and authority you give it. Between them sits a useful middle option: a fixed process with an AI step inside it.
Back at the hotel:
| The job | A sensible starting point | What it does |
|---|---|---|
| Save a conference enquiry and assign it to the events manager. | Workflow | Follows a rule using form fields. |
| Read a long email and extract the date, guest count and requested facilities. | AI inside a workflow | Interprets the message, then passes structured information to the next step. |
| Investigate comparable venue packages and resolve gaps in the brief. | Bounded research agent | Chooses sources, follows useful leads and asks for help when evidence runs out. |
More steps do not automatically mean you need more autonomy. If the hotel already has an approved brochure and a clear pricing rule, sending the right brochure does not require an agent to reinvent hospitality.
Anthropic's engineering guidance distinguishes predefined workflows from model-directed tool use and recommends starting with the simplest adequate approach.
Google also provides useful guidance. Its Agent Development Kit documentation describes agents built from models, instructions and tools, and workflows that combine agents with code. Its terminology is not identical to Anthropic's. Ask what the system chooses and what the software fixes in advance, rather than judging by the label.
And yes, Google's GEAR programme offers agent training, technical resources and samples. GEAR is a place to learn and build; it is not itself an agent you hire to run your marketing.
Where real businesses report real value
These are company-reported results, not guaranteed returns for your business. What matters is connecting each number to the work that changed.
Klarna: a marketing production bill that actually got smaller
Klarna's 2025 annual report, in its discussion of productivity, estimates $10 million in 2024 savings from AI-powered marketing. It also reports a $7 million year-on-year reduction in external marketing supplier spending in the first quarter of 2024, attributing 37% of that decrease to AI. The work includes translation, production, CRM and social agencies.
This is evidence of commercial value from AI-assisted marketing operations. It does not establish that autonomous agents produced those savings, or that every dollar cut from supplier spending came from AI.
The lesson for a smaller business is to inspect a recurring production expense. If you regularly pay to adapt approved campaign material for different formats or markets, test whether AI can reduce that expense while maintaining quality. There must be a real bill or a real capacity constraint to improve.
Generating more material is easy to count. Establishing whether you needed it is the better question.
Convert: faster answers, after fixing what the numbers meant
Google Cloud's Convert Human Digital Business account describes an agency whose early AI experiments produced unreliable answers. Its eventual Verus system combined five specialist agents with BigQuery and Looker's shared metric definitions, covering areas including media performance, creative analysis and ecommerce.
Reported results include:
- 65% less report-delivery time
- 8.4 times faster complex analysis
- 32% lower operational costs
The public account does not provide a full investment calculation or controlled comparison.
The commercially interesting detail is the work before the answer: documenting what each metric meant so the agents could use consistent business rules.
For our hotel, “bookings” might mean confirmed events to one person and every submitted enquiry to another. An agent can make either number arrive faster. Fixing the definition is what makes the answer useful for deciding whether the campaign deserves another week's budget.
That is a stronger business case than “we added a chatbot to the dashboard”.
Salesforce: working leads that were sitting idle
Salesforce's lead-nurturing engineering account reports an internal deployment that generated over $100 million in pipeline, created more than 10,000 opportunities, and contributed to 1,500 closed deals in its SMB segment. Agents handled outreach, qualification, and scheduling around inbound demand.
The account also describes early failure rates of 50% to 60% from overload and duplicate engagements before the team introduced stronger coordination. Its solution combined AI generation with deterministic workflows.
That makes the story more useful than the headline alone. Capacity was leaking after lead generation; the system helped work that demand. Reliable queues and duplicate prevention mattered alongside the model.
Pipeline is potential business, not collected revenue. The article does not disclose the closed deals' revenue, full operating cost, or a comparison group showing incremental sales. Investigate the lead bottleneck; do not mistake pipeline for a promised return.
What those examples mean for your budget
| Business | Reported result | What to investigate in your own operation |
|---|---|---|
| Klarna | Marketing cost savings | Repeated production and supplier expenses. |
| Convert | Faster reporting and lower operational costs | Manual analysis, inconsistent definitions and late decisions. |
| Salesforce | Pipeline and closed-deal contribution | Valuable inbound leads waiting for qualification or follow-up. |
The pattern is a recognisable bottleneck attached to a measurable cost or opportunity. None requires giving software the entire marketing department on day one.
Where I would start in a marketing team
Again, back to our hotel manager's Monday. Which unfinished job is expensive enough to deserve attention?
If enquiries sit unanswered, begin with capture, assignment and reminders. Use AI where the message needs interpretation. Consider a more adaptive agent if it must gather missing information or choose an appropriate next action across several systems. Track response time, qualified enquiries and eventual bookings.
If reports arrive after the decision, automate collection and calculations first. Add AI analysis once campaign names, dates and conversion definitions agree. An agent becomes worth investigating when people need to ask changing questions and follow up on unexpected findings.
If content production is the bottleneck, try an approved source article, a voice guide, and a defined set of deliverables. That can be a workflow with AI drafting steps. Open-ended research and changing evidence may justify an agent later. Measure usable output and editing time, rather than drafts generated.
If prospect research eats the morning, a bounded agent can investigate selected businesses, save sources and prepare a brief. Keep research and drafting separate from sending. Judge prospect fit and factual accuracy before treating a larger contact list as progress.
The choice should follow the obstruction. With few enquiries, a stronger offer may matter more than faster processing.
Make the business case before you make the agent
Start with one unit of useful work: an approved campaign report, a qualified enquiry or a publishable content package. Record the current volume, time, quality and expense.
Then include the work that remains after automation. Someone still provides context, reviews exceptions and repairs failures. Include setup, software, usage and maintenance too.
Here is an illustrative calculation, not a client result or market price:
| Monthly enquiry-processing work | Before | Proposed pilot |
|---|---|---|
| Enquiries processed | 200 | 200 |
| Human time per enquiry | 12 minutes | 4 minutes, including review and corrections |
| Total human time | 40 hours | 13 hours 20 minutes |
| Time valued at an assumed KES 1,000/hour | KES 40,000 | About KES 13,333 |
| Assumed software, usage and maintenance | KES 0 additional | KES 8,000 |
| Total monthly economic cost | KES 40,000 | About KES 21,333 |
That suggests about KES 18,667 in monthly economic benefit, assuming acceptable quality and that the freed capacity is useful. With an illustrative KES 60,000 setup cost, simple payback is roughly 3.2 months.
But if checking and correcting each enquiry takes ten minutes, the same calculation leaves only about KES 6,667 of labour capacity freed. Add KES 8,000 in running costs and the proposed system is more expensive on this measure.
No dramatic robot malfunction required. The supervisor's time quietly ate the return.
If the employee remains on the same salary, the time value is capacity, not cash removed from payroll. Cash savings require an actual reduction in spending. Revenue improvement needs evidence of additional sales, and the relevant contribution margin should be weighed against the costs of securing them. Do not count the same saved time twice.
Give it a trial period with an actual finish line
Before commissioning a broad build, run one narrow pilot through enough real cases to include the awkward ones.
For the hotel, that includes an incomplete date, an unusual package request and a customer replying after someone has already followed up. The system should show what it captured, what it changed, and what it handed to a person. Check the record and the delivered message, not just its “Done” response.
Google's agent evaluation guidance covers evaluating responses and the sequence of tool actions. The business translation is straightforward: check both the answer and the actions taken to produce it.
Agree on acceptable quality and human review time before the pilot. Compare against the current process and a simpler automation where one is plausible. Then decide whether to expand, revise, or stop.
Bring Attention Lasso one recurring marketing job, an example of the current output and the cost of getting it done. We can define what a useful pilot should demonstrate before you commit to a bigger system.
Frequently asked questions
Have businesses actually automated marketing and sales beyond demos?
Yes. Convert reports faster marketing analysis and lower operational costs, while Salesforce reports lead-nurturing agents contributing to closed deals. The useful test is whether a system completes valuable work reliably. Ask for operating results, including costs and human involvement, rather than a polished demonstration.
Will an AI employee save time, or give me another thing to manage?
It saves time only if it removes more work than it creates in supervision. Count the minutes spent preparing inputs, checking results and fixing mistakes. If a report takes an hour to write manually and ten minutes to review after automation, you have freed up fifty minutes.
Can an agent do my outreach without sounding robotic?
Yes, it can help produce natural outreach, but it needs a relevant offer, accurate prospect research and examples of your voice. Personalisation should explain why the message matters to that business. Adding a first name to a generic pitch is mail merge wearing a nicer shirt.
How do I stop automation breaking my CRM or campaign attribution?
Start with read-only access and require approval for changes until the system passes realistic tests. Define your fields, campaign names and duplicate rules first. Keep a change history and test retries: a connection failure should not become three copies of the same lead.
Do I need an agent, or can Zapier, Make or n8n do the job?
A workflow may be enough if you can define the steps in advance. Add AI where a step needs interpretation; consider an agent where the system must choose what to do next. Zapier, Make and n8n are platform choices. The important question is what your setup needs to decide.
Can an AI agent replace a marketing employee or agency?
It can take over specific tasks and reduce the work you need to buy. Replacing a whole role is a bigger claim. List everything that role delivers, including decisions, approvals and customer relationships. Then compare the cost and quality of completing the entire job.
How much should an AI agent cost, and when will it pay back?
Its price should reflect the job, volume and support required. Ask for setup and ongoing costs separately. To estimate simple payback, divide setup cost by monthly benefit after running costs. Spending KES 60,000 to gain KES 20,000 monthly suggests three months, provided the benefit holds.
What is the best first marketing process to automate?
Start with a frequent, time-consuming task whose result is easy to check. Enquiry assignment, report preparation or adapting approved content are useful candidates. Pick the job already causing delays, then measure time saved and output quality. It may need a straightforward workflow rather than an agent.
A new colleague needs more than a good introduction
By Friday, our hypothetical hotel manager should have something concrete to judge: enquiries reaching the right person sooner, less report preparation, or more time for customers. She should also know what the system still needs from her.
That is a useful working relationship. Whether the software earns an “agent” badge can wait.
Tell us where marketing work keeps getting stuck. We will start with the job and the value of completing it.
Keep pulling the thread
