AI Marketing Agents Blog | BoastOS

Stop counting the things your marketing agents generate

2026-09-24T00:00:00+00:00

An agent produced forty social posts this week. Is that good?

I can’t tell without knowing how many were useful, how long someone spent editing them, whether they were published, and what happened afterward. A larger draft folder can look like progress while the marketing team gets busier cleaning it up.

As agents take on more of the work, we need a better way to judge them than counting outputs.

Separate execution from results

First, check whether the workflow completed the assignment. If the job was to prepare a sourced competitor brief by Monday morning, did it arrive on time with working links and accurate changes? If the job was to publish an approved post, did the correct version appear in the right place?

Then ask whether the work helped the business. Did a useful brief change a decision? Did the published post bring relevant visitors? Did an outreach sequence lead to conversations worth having?

Those questions belong together, but they measure different things. A workflow can execute correctly and still support a weak campaign idea. It can also produce a promising draft while failing to carry out the task you assigned.

Keep a small set of real examples

Anthropic’s January 2026 article on evaluating AI agents discusses evaluations that look at the agent’s behavior and the resulting state, with graders suited to the task. A marketing team can borrow that discipline without turning every campaign into a research project.

Keep examples of work you would approve, work you would reject, and work that needs a judgment call. Include awkward cases: an expired offer, a source that contradicts another source, a customer quote without permission to publish, a brief with a missing audience.

Run those cases again when you change the workflow or its instructions. A change that improves your favorite example can still break another important case. You want to catch that before a customer does.

Count the work people still do

For a content workflow, I’d begin with a short weekly record:

How many drafts were submitted and how many were accepted?
How much review and editing time did they require?
How many approved pieces were actually published?
How many needed a correction after publishing?
What did the published work contribute to the campaign’s goal?

Use your own starting point. If the team previously spent two hours preparing a newsletter and now spends forty minutes, that is useful evidence. If it saves writing time but adds more coordination elsewhere, include that too.

Avoid pretending that one good week proves a business outcome. A handful of clicks or one booked call can tell you something happened. It cannot tell you the workflow will reliably produce the same result next month.

The article's hypothetical two-hour versus forty-minute example, plotted on the same scale. These are not customer results. Include all human effort in your own comparison. View full size.

Make the weekly record comparable

Use a consistent definition for each measure. These are suggested measures, not BoastOS benchmark results.

Measure	Definition	What it can reveal
Acceptance rate	Accepted drafts divided by submitted drafts	Whether preparation meets the team’s standard
Review effort	Total checking and editing minutes, including rejected work	Human effort hidden by a large output count
Completion rate	Finished assignments divided by assignments due that week	Work that stalls between draft and delivery
Post-release corrections	Published items requiring a factual or material correction	Problems the review process missed
Campaign outcome	The campaign’s chosen measure, such as qualified conversations	Whether completed work supports a useful goal

Record the raw counts alongside percentages. When no work was submitted, mark acceptance rate as not applicable. Compare similar assignments and include setup, coordination, and exception handling when assessing time saved. Attribution to an agent needs more evidence than a rise in total campaign results.

Use execution quality and business usefulness together to decide what to change. View full size.

Review the exceptions together

The failures often tell you more than the averages. Look at the draft that required a rewrite, the task that stopped halfway through, and the report that nobody used. Ask what the agent knew, what it attempted, and where someone had to step in.

Sometimes the fix is a better source document. Sometimes it is a narrower task or an earlier approval. Sometimes the team should stop running that workflow because the output doesn’t help anyone.

That’s a healthy decision. The point of agentic marketing is to give the business more capacity for useful work. A dashboard full of activity is only interesting when you can trace it to something your team needed done.

Bring the outcome you want to improve to a JumpStart conversation. We use discovery to choose the first useful agent work; pricing covers the software, implementation, and support options.

Where human approval belongs in agentic marketing

2026-08-13T00:00:00+00:00

An approval button is only useful if the person clicking it can understand what they’re approving.

“Approve campaign” sounds efficient. Does it approve the copy, the audience, the sending schedule, the budget, or all of them? If an agent can change any of those after the click, the approval doesn’t mean very much.

That is where I think marketing teams should spend more time as they adopt agents. The question is practical: where does a decision become a commitment someone outside the company can see?

Review the action while you can still change it

Research can usually run with little interruption. An agent can collect public competitor pages, summarize themes, and prepare a report. The output stays inside the business until someone decides what to do with it.

Publishing is different. So is sending an email to a prospect or increasing a campaign budget. Those actions deserve a review point that shows the actual content, destination, and scope.

For an outreach workflow, I would want to see the recipients, the message, the reason each recipient belongs on the list, and the planned send time. If the agent changes the message or adds recipients, that should become a new review. A previous approval shouldn’t silently stretch to cover new work.

Place human review before publishing, sending, or spending. Changes to the scope need a new review. View full size.

Make the review packet smaller

Human review fails when it becomes a wall of text nobody has time to read. A useful packet helps the reviewer decide quickly without hiding the details.

Show the proposed action first. Put the claims and their sources beside it. Explain the uncertain parts. If the agent couldn’t confirm whether a company still offers a service, say that clearly instead of burying a guess inside confident copy.

The reviewer needs a way to send the work back with a specific correction. “Wrong audience; use local owners rather than national chains” is actionable. A rejection with no explanation teaches the workflow very little.

Specify what the approval covers

For an outreach draft, the approval record could include the message version, recipient list, source of each factual claim, scheduled send time, and any spending limit. This is a design checklist, not a claim that every tool enforces these controls automatically.

Before enabling sending, try three cases:

Approve one message, then change the text. The changed version should require review again.
Approve a recipient list, then add another address. The new recipient should remain outside the approved send.
Interrupt execution after a send. On retry, the workflow should check whether the message already went out.

If the system cannot enforce those boundaries, keep execution manual until it can. A visible approval button is not enough to establish what was authorized.

An annotated interface concept showing what the reviewer needs to see. This is not a BoastOS screenshot or a claim that every tool enforces these controls. View full size.

Earn autonomy one workflow at a time

Anthropic’s guidance on building effective agents, first published in December 2024, describes agents that use feedback from their environment and can pause for human input. It also emphasizes testing and guardrails before trusting autonomous work. Those are useful principles for marketing, where a plausible draft can still contain an expensive mistake.

I’d apply them at the workflow level. A weekly internal summary and a public product announcement don’t need the same controls. A team might eventually let the summary run unattended while keeping approval on every announcement.

Keep a record of what was proposed, what was approved, and what actually happened. When something goes wrong, that record lets you find the failure. Did the brief contain the wrong fact? Did the reviewer miss it? Did the execution differ from the approved version?

Measure the burden on the reviewer

If your team spends an hour untangling every proposal, the workflow still needs work. Track how often reviewers approve without edits, what they change, and how long the decision takes. Use that evidence to decide whether to simplify the assignment, improve the context, or adjust the review point.

The goal is a person who can make an informed decision at the moment it matters. Good automation should make that decision easier to see and easier to act on.

Both BoastOS purchase options include human review and approval features. With JumpStart, we help build the first workflows around the decisions your team needs to review.

Your marketing agents need the same brief

2026-07-22T00:00:00+00:00

The draft reads well. The product name is wrong, the offer expired last month, and the closing paragraph promises a service you don’t provide.

Most marketing teams would call that a briefing problem if a freelancer delivered it. With AI, we tend to call it a prompting problem and start rewriting the instructions. Sometimes that helps. Sometimes we’re asking the model to compensate for a business that hasn’t agreed on its own source material.

An agent doing marketing work needs more than a tone of voice. It needs to know what is true today.

A shared brief needs an owner

In its September 2025 article on context engineering, Anthropic describes the work of selecting and maintaining the information available to an agent. The practical marketing implication is straightforward: what the agent can retrieve deserves as much attention as the instruction you type.

Start with the material a new person would need to work competently. What do you sell? Who buys it? What problems do customers bring to you? Which claims can you support? Which examples sound like you? Where can someone verify the current price?

Then assign ownership. A pricing document without an owner will eventually disagree with the website. A brand guide that nobody updates becomes a record of what the company used to sound like. Connecting those documents to AI makes the contradictions easier to distribute.

Give context an expiration habit

A service business might have an offer that ends on July 31. Put the end date alongside the offer, not in a separate email that the agent will never see. If an agent drafts an August promotion, it should be able to see that the offer is no longer available.

The same principle applies to customer stories, partner names, and product capabilities. Keep the source and review date close to the claim. Decide what the agent should do when two sources disagree. “Ask the owner” is a perfectly useful behavior when the alternative is publishing the wrong price.

This doesn’t require a giant knowledge project before you begin. For one workflow, a small set of current, approved documents may be enough. Expand it when the work calls for more.

Keep a maintained source set, then select only the context each assignment needs. View full size.

Keep a small context register

A shared folder is easier to maintain when each important document carries a few useful fields. Here is an illustrative register for a newsletter workflow:

Material	Owner	Review trigger	If it is stale or missing
Current offer	Campaign owner	Offer changes or expires	Omit the offer and ask for an approved replacement
Product facts	Product owner	A capability or price changes	Flag the disputed claim for review
Audience brief	Marketing lead	The campaign targets a new audience	Confirm the audience before drafting
Writing examples	Editor	Reviewers repeatedly make the same correction	Update the example and explain the change

Keep a link to the approved source with each record. Access should also match the assignment: customer information that an agent does not need should stay out of its context.

Don’t give every agent every document

A newsletter agent needs approved messaging, the audience brief, and the current offer. It probably doesn’t need every sales call transcript or an entire archive of abandoned positioning ideas.

That distinction also helps the person reviewing the work. When a draft makes a claim, ask the agent to point to the source it used. The reviewer can check the claim directly instead of guessing where it came from. If the source is stale, fix the source as well as the draft.

Over time, those corrections become a practical maintenance routine. You stop fixing the same misunderstanding in five separate conversations.

Look for fewer repeated corrections

The payoff isn’t that every sentence sounds identical. Different channels need different writing. The payoff is that the facts, promises, and business priorities stay consistent even when different agents do the work.

Watch the edits your team makes. If reviewers keep replacing the same phrase, correcting the same audience assumption, or removing the same unsupported claim, the shared brief needs attention. A better prompt in one chat only fixes one chat.

For us, that is the reason to put business context inside the marketing operation. The next agent should benefit from what the last review taught you.

Organizing that context is part of BoastOS JumpStart. You can also start with the software and training; the pricing comparison explains both options.

Your first marketing agent needs a job, not a job title

2026-06-11T00:00:00+00:00

A marketing leader can watch an AI demo over breakfast and have a dozen ideas before the coffee gets cold. By lunch, the question is harder: which one should actually run in the business?

Google’s May 20 announcements at Google Marketing Live 2026 put agentic technology directly into the marketing conversation. Its announced Ask Advisor agent spans Google Ads, Analytics, Merchant Center, and Google Marketing Platform. The direction is clear: software is moving further into work that used to require someone to interpret a screen and decide the next step.

For a small marketing team, that creates a useful opening. It also creates a temptation to start with an enormous assignment: “Run our marketing.”

I’d start with something much smaller.

Give the agent a finish line

“Content strategist” sounds like a sensible role. It doesn’t tell an agent when the work is done.

Try this instead: every Tuesday, review the questions customers asked last week, identify one we haven’t answered on our website, and prepare a draft for review. Include the source questions, the intended reader, and the existing pages you checked.

Now there is an input, a cadence, and an output someone can judge. There is also a clear boundary. The agent prepares the work; a person decides whether it is ready to publish.

This distinction matters because a capable model can make a weak assignment look productive. It will produce outlines, plans, and suggestions. Someone still has to turn those suggestions into a dependable process. If that person is you, count the time you spend doing it.

A bounded first workflow: prepare a sourced draft, then hand the decision to a person. View full size.

A brief you can reuse

Before you build, complete these fields together. This is an example assignment, not a report of a customer deployment.

Field	Example
Job	Prepare one draft answering a recurring customer question
Input	Last week’s approved, anonymized customer questions and current website pages
Cadence	Tuesday morning
Finished output	A draft with source questions, intended reader, and links to related pages
Boundary	Do not publish or invent a price, offer, or customer quote
Reviewer	The marketer responsible for the website
Stop condition	Ask the reviewer if the sources disagree or required information is missing

Run a few examples before scheduling it. Include one where there is no worthwhile new question. Returning that finding can be more useful than manufacturing a topic.

Pick work you already understand

Your first workflow should be familiar enough that you can spot a bad result. Competitor monitoring is a good example when your team already knows the competitors. The agent can check a defined set of pages, compare what changed, and bring back a short brief with links. You can tell whether the change matters.

An unfamiliar channel is a harder first project. You would be evaluating the channel strategy and the automation at the same time. When the result disappoints, you won’t know which part needs attention.

Write down what a useful output looks like before you configure the agent. Keep a couple of examples. Record the mistakes that would make the result unusable. A beautiful report about the wrong competitor is still the wrong report.

Keep the first handoff visible

Imagine a plumbing business that wants a monthly email for existing customers. The first version could collect seasonal maintenance questions, propose a topic, and draft the email. The owner checks the advice and approves the offer. Sending can become a separate step once the team trusts the preparation.

That’s enough to learn something. Did the draft use the services the business actually sells? Did it invent a discount? Did the owner spend less time getting a useful email ready?

You can add another workflow after the first one works reliably. You can also decide that an ordinary template solves part of the problem more cheaply. Both are good outcomes.

The test I would use after a month is simple: does a recurring piece of marketing work now get finished with less chasing? If it does, you’ve earned the next step. If it doesn’t, giving the agent a grander title won’t fix it.

If choosing and building that first workflow keeps getting pushed aside, BoastOS JumpStart starts with your business and the work worth automating. Compare what is included on the pricing page.

Your SaaS just became plumbing: build for agent access

2026-05-06T00:00:00+00:00

A customer wants a campaign report. Today, they may open three applications, export data, and assemble the answer. An agent could do that work through the same applications’ APIs, then bring the report back for review.

The applications still matter. Their interface may no longer be where the customer starts.

That is the strategic shift I want SaaS builders to consider. Adding an AI assistant inside your product is one response. Making your product dependable when another agent calls it deserves equal attention.

What the orchestration layer does

An orchestration layer coordinates steps, tools, and saved state. A model may choose the next action; software provides the tools, permissions, and record of what happened. In a marketing workflow, that could mean retrieving campaign data, comparing it with a brief, drafting a recommendation, and pausing for approval.

LangGraph’s documentation describes infrastructure for stateful agent workflows, including durable execution and human intervention. That is a concrete example of the layer. It is not evidence that every business has adopted it or that every workflow needs it.

Anthropic’s guidance on building effective agents distinguishes predefined workflows from agents that choose their own steps. It recommends simpler approaches when they are sufficient. A scheduled API call may solve your customer’s problem without an autonomous agent.

Why this matters to a SaaS business

When a customer can complete a task without opening your interface, some of your value becomes less visible. The report may appear in a chat or another application’s workspace. Your product still supplies data or performs an action, but the customer’s attention is elsewhere.

My view is that this makes reliable integration more important to retention and product strategy. It does not make interfaces, branding, or onboarding worthless. People still need to configure systems, investigate mistakes, resolve exceptions, and understand what they are paying for.

Consider a marketing contact database. An agent might retrieve approved contacts and draft outreach. The database still needs accurate records, consent information, access controls, and a usable way for people to correct mistakes. An agent does not remove those responsibilities.

Model choice does not remove switching costs

An orchestration layer can make it easier to route different tasks to different model providers. That gives buyers options. It does not make providers interchangeable.

Tool use, response formats, cost, speed, and error patterns can change between models. A workflow that performs well with one model needs evaluation before a replacement handles customer data or external actions. Credentials and data-handling requirements also remain part of the decision.

For a SaaS builder, the practical opportunity is to keep the integration contract clear enough that a customer can change their agent setup without rebuilding your side of the connection. You cannot control every model they choose. You can document what your API accepts and what it returns.

Give agents a supported way to act

Agents can use APIs, tool protocols, command-line utilities, or browser automation when the customer has authorized that access. Those routes have different failure modes. An API with stable identifiers is usually easier to validate than a sequence of clicks that depends on a screen layout.

MCP is one option for exposing tools and resources. Its authorization specification describes an authorization framework for HTTP-based transports. Using MCP does not by itself decide which business actions an agent should be allowed to perform. Your application still needs to enforce its permissions.

Browser access is also not permission to bypass authentication, contractual restrictions, or approval requirements. If your customer’s agent cannot legitimately complete a task, provide an explicit failure or a supported handoff.

Start with one customer task

Pick a task a customer already repeats. For example, retrieve last week’s campaign results and prepare a report. Test the complete sequence, including a missing campaign, an expired credential, and a request repeated after a timeout.

Use these checks to review your integration:

Question	Evidence to ask for
Can the agent find the right record?	Stable IDs, documented filters, and clear pagination
Can access be limited to the task?	A read-only credential for a reporting workflow
What happens after a timeout?	A retry policy and protection against duplicate writes
Can a person reconstruct the action?	Logs that identify the actor, inputs, result, and approval where required
What happens when the data is missing?	A clear error or incomplete result, rather than an invented answer
Can the customer stop the workflow?	Credential revocation and a documented way to cancel work

These are product decisions as well as engineering decisions. A report that quietly omits half the requested data is a customer problem even if the API returned a successful status.

Revisit pricing with actual usage

Agents can change the relationship between human seats and the amount of work performed. One employee might initiate many automated jobs. That is a reason to measure usage and support costs before deciding how to charge.

It is not proof that every seat-based plan must disappear. A seat can still represent governance, collaboration, or access to expert tooling. Usage-based pricing also has tradeoffs: customers need predictable bills, understandable units, and limits that keep an accidental loop from becoming an expensive surprise.

Review a real workflow with a customer. Record the calls, compute, data access, and support it requires. Use that evidence to decide whether your current plan makes sense.

Build for the work, wherever it starts

The useful question for a SaaS leadership team is: if a customer asks an agent to use our product, can it complete an authorized task reliably, and can a person understand the result?

Answer that with a working integration and a small set of tests. Keep the human interface useful for oversight and exceptions. Make the supported actions clear enough that customers do not have to improvise around your product.

That is the part of this shift you can act on now. For marketing teams choosing the work to automate, our guide to a first agent assignment starts at the other end of the same problem.

For a marketing workspace built around agent tasks and human review, explore BoastOS pricing and installation. If you need help choosing the first workflow, JumpStart includes discovery and implementation.

Welcome to BoastOS

2026-05-04T00:00:00+00:00

A useful AI draft can still leave a marketer with most of the job: find the right facts, move the work between tools, check the output, get approval, and make sure something actually ships.

We are building BoastOS to help teams manage that work. It brings agents, business context, tasks, schedules, and review into one marketing workspace. Our aim is for the person responsible for the work to understand what an agent is doing and decide what happens next.

What stays with the marketing team

Your team defines the audience, maintains the business facts, and decides which actions need approval. An agent can help prepare and carry out work, but it still needs a clear assignment and access to the appropriate tools.

That applies whether one owner does all the marketing or a department shares the responsibility. The size of the team changes the handoffs. It does not remove the need for accurate information and someone accountable for the result.

Start with a job you can judge

A sensible first assignment might be a weekly competitor brief: review a defined set of public pages, identify changes, and return a short report with links. Someone on the team checks whether the findings are accurate and useful before relying on them.

The test is practical. Does the brief arrive when needed? Can the reviewer verify it? Does preparing it take less time once checking and corrections are included?

Some workflows need technical setup, account permissions, or specialist review. We want to make routine operation accessible to marketers while being clear about those dependencies. JumpStart pairs the BoastOS license with hands-on discovery, setup, and guidance for the people who will run it.

What we will write about

This blog covers agent assignments, shared context, approvals, and ways to measure completed work. We will distinguish illustrative examples from measured results and explain the limits of evidence we cite.

If you are deciding where to begin, read how to give your first marketing agent a clear job. Bring a recurring task your team understands, a useful example of the finished work, and the name of the person who will review it.

For installation, training, and support details, compare the BoastOS purchase options.