From 10 AI Tools to 3: Deciding What to Cut

You are paying for a stack of about ten AI tools, and most of them never hand work to the next step. Each one arrived because someone had a real problem that week: scoring, a draft, a dashboard nobody trusted. Together they are a pile of logins and invoices, and the team still opens several of them to answer whether a campaign is working. The average B2B marketing organization now manages seven separate data sources that were never designed to work together, according to Salesforce's State of Marketing 2026, which surveyed 4,450 marketing professionals. Only 8% of marketing teams orchestrate multi-step AI workflows, according to NinjaCat's 2026 AI Maturity in Marketing report (n=500+). This is how you reduce a marketing AI tool stack to the few tools that earn a place in a running workflow.

Teams collect AI tools one pain at a time, and the pains never share a workflow.

The stack grows in a pattern you can recognize after the third renewal. A tool gets purchased for one pain: lead scoring, content generation, or campaign reporting. It works well enough alone. Then another tool arrives for a different pain. The two never pass a lead between them. A third tool shows up because someone is tired of reconciling the first two by hand. Each addition brings a login, a dashboard, a vendor review, and a monthly invoice. None of them is large enough, on its own, to kill in a budget meeting. All of them together are large enough to slow the team down.

The purchase is usually made by the person who feels the pain.

Most of these purchases are made by the practitioner who needs the capability that week, rather than by someone who is designing the workflow. That is what a fast market does. Individual features ship faster than a buying process can evaluate them. The result is a set of tools that each solved their original problem and were never built to hand off a contact, share a field, or start the next step. They sit in parallel. A person reconciles them at the end of the week and calls that the process.

The stack outgrows the team's ability to run it.

NinjaCat's 2026 report (n=500+) found that 89% of marketing teams use three or more separate tools just to identify performance issues. When it takes three tools to answer whether a campaign is working, the stack is larger than the team can orchestrate. The same report found that 91% of marketing teams say AI changed some part of their work for the better, and that 72% still rely on manual reporting. The two figures describe the current state cleanly. The tools are running. The outputs still land in a manual reconciliation, rather than in the next automated step.

A workflow tool moves the work, and a reporting tool waits for a person.

Cost is a weak first cut, because a cheap tool inside the sequence is worth more than an expensive tool that produces a weekly PDF. The useful lens is function. Does this tool sit inside a sequence of automated steps, or does it produce an output that a person has to read before anything happens?

One seat receives work and passes it on.

A workflow tool receives an input from the previous step, does something with it, and passes an output to the next step. An email personalization engine that fires when a lead crosses a score threshold is a workflow tool. A lead scoring dashboard that tells you which leads to call is not. The dashboard produces information. The workflow tool moves the work without requiring someone to read the output and then decide to act. Monitoring still has a job. Real-time analytics, anomaly detection, and campaign reporting are legitimate. They do not earn the same renewal decision as a tool inside the sequence. Name the seat in plain words before you decide to keep the invoice.

When you label the seat, write the next action the tool causes. If the next action is that someone opens a dashboard, it is a report. If the next action is that the CRM updates and the sequence continues without that person, the tool is in the workflow. Do this for every invoice before you sort by price. A tool a person likes to open by hand can stay on that person's desk. It should not survive the same renewal test you use for a tool that moves a lead.

Orchestration only works on tools that hand something off.

NinjaCat found that 8% of teams run multi-step workflows. A tool that hands off nothing cannot be part of that kind of workflow, however capable it is alone, which is the argument in why AI tool integration fails. That is why the cut starts with orchestration fit. Price comes after you know whether the tool is even in the sequence.

Three questions tell you whether an AI tool earns its seat.

Answer them in order. A tool that fails the first question does not get a long debate on the second.

The first question is whether the tool is inside a workflow that runs today.

Pull the workflows that run without someone starting them by hand: lead score updates, email triggers, CRM field updates, routing rules, and alerts. Map each AI tool against that list. If the tool appears inside at least one sequence, it is a workflow participant. If it appears in none, it is a monitoring or reporting tool, and its value has to be justified on those terms. A tool in zero running workflows is a cut candidate, whatever the invoice says.

The second question is whether you are paying for the same job twice.

Before you ask whether a tool is good, ask whether another tool already covers the job inside a workflow you actually run. Email personalization, lead enrichment, campaign analytics, and content generation each show up across multiple tools in a stack that grew by purchase. You do not need the strongest tool in a category if another tool already covers that category well enough inside the sequence.

A 15-minute pass finds the overlap.

List every tool. For each one, write its primary output in one sentence. Then look for outputs that describe the same job. "Identifies high-intent leads from web behavior" and "surfaces high-intent accounts from website visits" are the same ground, sold twice. You need one of them. Run this pass before you negotiate a renewal, because a renewal conversation that starts from price skips the overlap and keeps both.

The third question is whether you can name what it changes.

McKinsey's State of AI 2025, which surveyed 1,363 respondents, found that only 19% of organizations using gen AI track gen AI-specific KPIs, and that over 80% are not seeing enterprise-level EBIT impact from their gen AI investments. If you cannot name the metric a tool moves and the baseline it started from, the tool is occupying budget without anyone owning the outcome. That is the cut, even when the team likes the interface.

You can run the cut in a single afternoon.

The three questions become a spreadsheet that takes three to four hours with your revenue-operations or marketing-operations lead. You do not need a committee or a vendor bake-off. You need one person, the tool list, the active workflows, and three columns.

Three columns are enough to see the cut list.

Column A is the tool name, the vendor, the monthly cost, and the contract end date. Column B is workflow participation: yes or no, and which workflow. Column C is the primary output in one sentence, plus the business metric it claims to move and your current reading of that metric. Fill every row. Rows where column B is blank, and column C does not name a metric that is currently moving, are the cut list. If a tool has no workflow seat and no measured impact, the burden of explaining the renewal sits with the person who wants to keep it.

Sort by whether it is in a workflow, then by cost.

Tools with workflow participation and a measured impact go to the top. Tools with no workflow participation go to the bottom. Anything in the bottom half that has been live for more than 90 days without a measurable workflow role is a cut. The 90-day mark exists so "we just started using it" cannot defend a tool that has renewed quietly for two years with no defined job. Salesforce's 2026 research (n=4,450) found that 87% of marketing teams use gen AI in at least one recurring workflow, and that only 13% have deployed agentic AI. Most stacks already have some workflow underneath them. What they lack is the willingness to cut the tools that are not in it.

A lean stack covers three jobs, with one primary tool in each.

There is no universal trio of product names. There is a useful structure. A lean AI stack for a B2B marketing team covers three jobs. Each job needs at most one primary tool with full workflow participation. A second tool in the same job is overlap or a leftover, not a strategy.

Scoring, delivery, and measurement are the three jobs.

The first job is intent and scoring: the tool that identifies which contacts and accounts are moving toward a buying decision. It feeds the CRM and it triggers follow-up. It is in the workflow from the day you keep it. The second job is orchestration and delivery: the tool that runs the sequences scoring starts. Email, SMS, in-app, and social delivery happen here. It receives from the first job and passes completion signals to the third. The third job is attribution and measurement: the tool that reads the CRM contact and opportunity fields and reports which activities influenced pipeline and revenue. It does not run the workflow. It reads what the workflow wrote.

When each purchase happens without this structure, these three jobs get obscured, which is why the build-or-buy guide starts every purchase from the workflow. Most bloated stacks have two or three tools competing for scoring, one or two tools that partly cover delivery, and three or four tools that each claim measurement with methods that do not agree.

What leaves is whatever sits outside the sequence.

A content tool that never connects to delivery is a cut candidate. An intent feed that never reaches the scoring model is a cut candidate. An analytics platform that does not read the same CRM fields the measurement job uses is a cut candidate. The tools that stay are the ones inside the sequence. Why AI tools sit in isolation covers the mechanical reasons a disconnected tool fails to compound with the others.

A tool you cannot exit yet still needs an owner and an end date.

Annual contracts are the default, and most of these purchases were made before anyone had this filter. A tool that fails the cut may be locked for another eight months. You do not wait eight months to decide. You make the remaining term earn a place, or you make the non-renewal someone's job.

Write the end date, the owner, and one attempt to wire it in.

For each locked tool that failed the cut, write three things. The end date is the contract end. The owner is the person responsible for not renewing. The task is one concrete attempt to wire the tool into a workflow before that date. The task matters because some tools earn a seat once someone finally connects them. If the attempt fails and the tool is still outside every workflow when the contract ends, do not renew. Write the decision in the spreadsheet so the next budget review can see why it was dropped.

A replacement and a leftover should not both receive new work.

When a replacement is being onboarded while the old tool is still under contract, run both for a defined window and then stop. Set a date 30 to 60 days from the start of onboarding when the old tool stops receiving new data. After that date it is read-only, for history. It does not participate in new workflows. Leaving both active "in case we need it" is how a dead tool survives three renewals.

Consolidation pays off after a short dip you can plan for.

The dip is real, and it lasts four to eight weeks. A team that cuts from ten tools to three loses the habits embedded in the old interfaces, the workarounds around their quirks, and the comfort of familiar dashboards. Plan for that dip. It is shorter than the time the team is already losing to switching tools and reconciling three dashboards every Monday.

Most of the dip is rewiring, and it is front-loaded.

The first three weeks carry most of it. Sequences that depended on cut tools have to be rebuilt in the tools you kept. Plan about four hours of revenue-operations time per workflow that needs rewiring. A team with five active workflows and two cut tools typically needs 10 to 20 hours of rebuild time, spread across two to three weeks. That is the cost of the dip. After the rebuild, the team has fewer logins, fewer reconciliation steps, and fewer Monday questions about which dashboard to believe. Error handling routes for AI marketing workflows is the rebuild pattern for when a tool leaves a sequence and the failure path has to move with it.

Three signals tell you the cut worked.

First, the team can answer whether a campaign is working from a single workflow output, rather than by opening three dashboards. Second, a new contact moves from capture through scoring to delivery to attribution without a person pushing it between steps. Third, a new teammate can learn the stack in under two hours, because there are three tools to learn. Get your free Systems Plan before the cut if you want a written baseline across the workflow dimensions you already measure, and read the same dimensions after the tools are gone. The change between those two reads is what the decision returned.

The methodology behind this cut.

This post sits in the workflow orchestration series. The main guide is Choosing Your First AI Workflow. The three questions (a seat in a running workflow, overlap with a tool you already have, and a metric you can name) apply that guide to the problem of a stack that grew by purchase. The external figures are from NinjaCat's 2026 AI Maturity in Marketing report (n=500+ marketing professionals): 8% orchestrate multi-step AI workflows, 89% use three or more tools to identify performance issues, 91% say AI changed some part of their work for the better, and 72% still rely on manual reporting. Salesforce's State of Marketing 2026 (n=4,450 marketing professionals, fielded October to November 2025) is the source for seven separate data sources, for 87% using gen AI in at least one recurring workflow, and for 13% having deployed agentic AI. McKinsey's State of AI 2025 (n=1,363) is the source for 19% tracking gen AI-specific KPIs and for over 80% not seeing enterprise-level EBIT impact.