# AI Tools for Marketing Teams: What to Buy, What to Build, What to Skip > A research-backed way to decide which AI tools your marketing team should buy, configure, build or skip, with a four-question sorting tree, a worked example and team rules that beat bans. - **Author**: Marxx - **Published**: 2026-10-24 - **Category**: Agency operations - **URL**: https://marxx.ai/posts/ai-tools-for-marketing-teams --- **Which AI tools should a marketing team buy, build or skip?** Buy tools that bring data or infrastructure you can't get yourself, such as competitor ad intelligence or ad production at volume. Configure, rather than build, anything a saved prompt, an official connector and your own documents can do. Build only what's core to how you win and what you can maintain. Skip anything that doesn't replace a weekly job. ## The problem isn't too few AI tools. It's no way to choose. Every week a new AI tool promises to save your marketing team hours. Teams usually react in one of two ways: buy everything, or ban everything. The evidence suggests both go wrong. **Buying everything** fills the stack faster than people can use it. Gartner found that marketing teams used only **33%** of their martech stack's capabilities in 2023, down from 58% in 2020 ([Gartner, 2023](https://www.gartner.com/en/newsroom/press-releases/2023-08-23-gartner-survey-finds-63-percent-of-marketing-leaders-plan-to-invest-in-generative-ai-in-the-next-24-months)). **Banning everything** doesn't stop use; it just hides it. In Microsoft and LinkedIn's 2024 survey of 31,000 people, **75%** of knowledge workers used AI at work, and **78%** of them brought their own tools ([Microsoft, 2024](https://blogs.microsoft.com/blog/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work/)). And even deliberate investments often miss. MIT's NANDA initiative studied enterprise generative AI in 2025, using 150 leader interviews, a survey of 350 employees and 300 public deployments. It found that only about **5%** of pilots produced rapid revenue gains ([Fortune, 2025](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/)). So the question isn't "should we use AI?" It's "**how do we decide, tool by tool, what to buy, what to make ourselves, and what to ignore?**" This guide gives you: 1. What the research says about buying versus building. 2. A four-question sorting tree, with a fourth bin most teams miss. 3. The biases that push teams towards bad decisions. 4. A worked example, team rules, and a quarterly review process. Three bins: buy when the job needs data you don't own, build when it's a prompt plus your own data, skip when it saves less than an hour a week --- ## Part 1: What the evidence says about buying versus building MIT NANDA 2025: about 5% of AI pilots delivered rapid revenue gains; buying from specialised vendors succeeded about 67% of the time; internal builds succeeded about one-third as often The MIT NANDA findings, as reported by Fortune, carry two lessons for marketing teams. **Lesson 1: Bought tools succeeded more often than internal builds.** Purchases from specialised vendors and partnerships succeeded about **67%** of the time. Internal builds succeeded roughly **one-third as often**. **Why that might be:** A specialised vendor has already solved the hard, unglamorous parts: data pipelines, edge cases, maintenance and updates. An internal team usually starts from zero and underestimates the work after launch. **Lesson 2: Marketing gets the budget, but not necessarily the return.** More than half of generative AI budgets went to sales and marketing tools. The biggest returns showed up in back-office automation, such as cutting outsourcing and agency costs and streamlining operations. Where budgets go versus where ROI showed up: more than half of gen AI budgets to sales and marketing tools; the biggest returns in back-office automation **What this means for you:** The flashy tools, like a new image generator, aren't necessarily where the payback is. The boring jobs often pay back first: the weekly report, the reformatting, the brief nobody wants to write. **Limitations:** This is a report about large enterprises, summarised by the press, with "success" defined by the researchers. It isn't a controlled study of marketing teams. Treat it as a strong signal about where projects tend to stall, not as a law. > **The quotable version:** Marketing gets more than half the AI budget. The boring jobs are where the return showed up. --- ## Part 2: The sorting tree The original "buy, build, skip" idea is useful but missing a bin. Most things marketing teams call "building an AI tool" aren't builds at all. They're **configuration**: a saved prompt, an official connector and the documents you already have. Most build ideas are really configure: configure means a saved prompt, an official connector and your documents, set up in hours by a marketer; build means custom code, integrations, testing and upkeep, owned by an engineer That distinction matters because the MIT finding is about *engineering* builds. Configuration is a different thing: cheap, fast, owned by marketers, and easy to throw away. ### The four questions The sorting tree: does it replace a weekly job, if not skip; does it need data or infrastructure you can't get, if so buy; can a prompt plus official connector plus your docs do it, if so configure; is it core with engineers to maintain it, if so build, otherwise buy Ask them in order. Stop at the first "yes" that sorts the tool. **1. Does it replace a job someone does every week?** If not, **skip**. A tool without a recurring job is a demo, not a tool. This question alone cuts most wish lists in half. **2. Does it need data or infrastructure you can't get yourself?** If so, **buy**. Examples: thousands of competitor ads collected and tagged every week; voice, avatar and video assembly at scale. These are expensive to build *and* to keep running. **3. Can a saved prompt, an official connector and your own documents do it?** If so, **configure**. Examples: a weekly results summary through an MCP server; brand voice rules in your assistant's project instructions; a Slack digest. **4. Is it core to how you win, and do you have engineers to maintain it?** If both are true, **build**. Otherwise, **buy**. Most marketing teams answer "no" to the second half more often than they expect. **Where the tree misleads:** The answers change as tools improve. Something you'd have had to buy last year might be configurable this year. Re-run the sort every quarter. --- ## Part 3: The biases that push teams to the wrong bin ### The IKEA effect: Overvaluing what you built The IKEA effect: people value things more when they have built them themselves, but only if they finish them; judge AI builds by use, not pride Behavioural researchers Michael Norton, Daniel Mochon and Dan Ariely found that people place a higher value on things they've built themselves: the **IKEA effect**. The bonus disappeared when people failed to complete what they were building ([Norton, Mochon & Ariely, 2012](https://dash.harvard.edu/handle/1/12136084); [JSTOR Daily summary](https://daily.jstor.org/why-we-pay-to-do-stuff-ourselves/)). **How it shows up:** The team that spent a month wiring up a custom reporting bot will defend it long after a bought tool would have done the job better. And half-finished internal tools, the ones that never quite work, tend to linger. **The fix:** Judge every build by the same measure as a purchase: hours saved per week, and how many people actually use it. ### Demo bias: Judging tools by their best moment Demos show a tool at its peak, on clean data, run by an expert. Your Monday is messier. The fix is the same as for ad spy tools: test it on your own data, on a normal week, before you buy (see [our ad spy tool checklist](/blog/ad-spy-tool-checklist)). ### Sunk cost: Keeping tools because you paid for them Annual contracts make it painful to admit a tool isn't used. The question to ask isn't "what did we pay?" but "**would we buy this today?**" If not, plan to cancel at renewal. --- ## Part 4: A worked example *Hypothetical example: a six-person paid social team's wish list.* A team's wish list sorted: competitor ad tracking buy; weekly results summary configure; brand voice checker configure; ten ad versions a week buy; custom attribution model build maybe; fifth headline generator skip; AI meeting notes skip if the suite already has it; client Slack digests configure | Request | Q1: Weekly job? | Q2: Data you can't get? | Q3: Prompt + connector? | Bin | Why | |---|---|---|---|---|---| | Competitor ad tracking | Yes | Yes: thousands of ads, tagged | -- | **Buy** | Collecting and tagging at scale is a product, not a prompt | | Weekly results summary | Yes | No | Yes | **Configure** | An MCP connector to the ad account plus one saved prompt | | Brand voice checker | Yes | No | Yes | **Configure** | A one-page voice guide in the assistant's instructions | | Ten ad versions a week | Yes | Yes: voice, talent, video, resizing | -- | **Buy** | Production infrastructure isn't worth building | | Custom attribution model | Monthly | Partly | No | **Build, maybe** | Only if it's core and you have an analyst to own it | | A fifth headline generator | No new job | -- | -- | **Skip** | The assistants you already have do this | | AI meeting notes | Yes | No | -- | **Skip as a purchase** | Most video-call suites now include it | | Client Slack digests | Yes | No | Yes | **Configure** | Many tools, Marxx included, post digests to Slack | **What the sort revealed:** Of eight requests, only two needed new spending. Three were configuration jobs the team could finish in a week. One was a genuine build question that needed an owner before anything else. Marxx creative board grouping competitor ads by hook, angle and format ### What "configure" looks like in practice **The weekly report:** Connect your ad data to Claude or ChatGPT through an MCP server, then save one prompt: *"Summarise last week's results, the three biggest changes, what's fading, and what competitors launched. End with one thing to test next week."* Run it every Monday. Marxx answering a question about ads inside Claude **Brand voice:** Write one page on how your brand sounds and doesn't sound, with three good and three bad examples. Add it to your assistant's project or instructions. You don't need a separate "brand voice AI". **Team digests:** Not everyone needs to open an AI assistant. Marxx posts competitor briefs and weekly digests to Slack, so the whole team gets the insight without another login. A Marxx competitor brief delivered in Slack --- ## Part 5: Rules, not bans With 78% of AI users bringing their own tools, banning AI mostly means losing sight of how it's used. A short set of rules works better. Team rules: approved list, data rule, connector rule, write rule, quarterly review 1. **Approved list:** One page stating which tools are used for which jobs. 2. **Data rule:** No customer or client data in personal AI accounts. 3. **Connector rule:** Official or vetted MCP servers only. Unknown servers can carry hidden instructions (see [best MCP servers for marketing](/blog/best-mcp-servers-marketing)). 4. **Write rule:** Anything that spends money or changes a live campaign needs a human yes. 5. **Quarterly review:** Re-sort every tool. Cancel anything nobody opened last month. ### How to measure whether the stack is working | Metric that matters | Why | |---|---| | Weekly active users per tool | Shows real use, not seats paid for | | Hours saved per job per week | The actual return | | Cost per active user | Exposes expensive tools few people use | | Saved prompts reused each week | Shows configuration is working | | Metric that misleads | Why | |---|---| | Number of AI tools | More tools often means lower use of each | | Seats purchased | Paid isn't used | | Vendor-reported "time saved" | Measured on their best case, not your week | --- ## Six mistakes, and how to spot them | Mistake | Symptom | Fix | |---|---|---| | Buying for a demo | A tool nobody opened last month | Ask question 1: what weekly job does it replace? | | Building what you could configure | A month of engineering for a saved prompt | Ask question 3 before question 4 | | Building without an owner | A half-finished internal tool | No owner, no build | | Banning AI | Staff use personal accounts anyway | Rules, not bans | | Spending only on flashy tools | Great demos, same workload | Start with the boring weekly jobs | | Never re-sorting | Paying for last year's needs | Quarterly review | ## What to take into your next planning meeting - **Add a fourth bin.** Most "build" ideas are really "configure": a saved prompt, an official connector and your own documents. - **Bought tools succeeded more often.** In MIT NANDA's research, vendor purchases succeeded about twice as often as internal builds. - **The boring jobs pay back first.** Marketing gets more than half of AI budgets, but the returns showed up in operations. - **People overvalue what they build.** Judge builds by use, not effort. - **Rules beat bans.** Three in four knowledge workers already use AI, and most bring their own. ## FAQ **What AI tools should a marketing team buy?** Tools that bring data or infrastructure you can't build yourself, such as competitor ad intelligence, ad production at volume, and official connectors to your ad and analytics accounts. **Should a marketing team build its own AI tools?** Usually not from scratch. Most needs can be met by configuring an assistant with a saved prompt, an official connector and your documents. Build only what's core to how you win and what you can maintain. **How do we stop AI tool sprawl?** Sort every tool with the four questions each quarter, measure weekly active users, and cancel anything nobody opened in the last month. **Is it safe to let staff use their own AI tools?** Not without rules. Set an approved list, keep customer data out of personal accounts, and use only official or vetted connectors. --- ## Add the Marxx MCP server to Claude or ChatGPT The "configure" bin starts here: connect your ads and your competitors' ads to the assistant your team already uses, save the prompts you run every week, and send digests to Slack. **[Add the Marxx MCP Server ->](https://www.marxx.ai)** *Related: [Best AI Tools for Marketers, by Job](/blog/best-ai-tools-for-marketers) - [Best MCP Servers for Marketing](/blog/best-mcp-servers-marketing) - [Guardrails for Letting AI Touch Your Ad Account](/blog/ai-ad-account-guardrails)* ### Sources - Gartner (2023). [Marketing leaders and generative AI; martech utilisation](https://www.gartner.com/en/newsroom/press-releases/2023-08-23-gartner-survey-finds-63-percent-of-marketing-leaders-plan-to-invest-in-generative-ai-in-the-next-24-months). - Microsoft and LinkedIn (2024). [2024 Work Trend Index](https://blogs.microsoft.com/blog/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work/). - Fortune (2025). [MIT report: 95% of generative AI pilots at companies are failing](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) (MIT NANDA, *The GenAI Divide*). - Norton, M., Mochon, D. & Ariely, D. (2012). [The IKEA effect: When labor leads to love](https://dash.harvard.edu/handle/1/12136084). *Journal of Consumer Psychology*, 22(3). --- - [More Agency operations articles](https://marxx.ai/posts/category/agency-operations) - [All articles](https://marxx.ai/posts)