Which AI tools does an agency need? Fewer than most people think – but better connected. The bottleneck is rarely a missing tool; it is missing integration. The data sits in five systems, every team has its favourite chatbot, and at the end of it someone pastes screenshots into a deck.
The big agency groups, meanwhile, are betting on connection over variety. According to Digiday, Dentsu's media agency Carat is developing an agent for planning, buying and optimisation in the upfronts with NBCUniversal, FreeWheel and AI company Newton. Tinuiti has launched an MCP server for its Bliss Point measurement solution, opening up its measurement layer to AI agents and bringing audience, creative, media and measurement data under one roof. And WPP Media stresses that ‘governance and human accountability are central to agentic media buying’.
Why do more AI tools often deliver less?
Because every new tool creates a new island. The social listening tool knows the conversations but not the media data. The chatbot writes brilliantly but knows nothing about the client. The reporting tool has numbers but no context. A human makes the connection between them all – by copy and paste.
Then come the hidden costs. Overlapping licences. Onboarding for every dashboard. Shadow AI: personal accounts through which client data travels without anyone having approved it. And knowledge that depends on the one person who knows which prompt works in which tool.
For agencies in Germany, whose clients look hard at data protection and data sovereignty, this sprawl isn't just inefficient. It is a risk that gets exposed in the next pitch questionnaire.
Why the whole is more than the sum of the tools
Systems theory offers a simple explanation: a system is more than the sum of its parts, because its performance comes from the relationships between them. An orchestra of brilliant soloists without a shared score doesn't sound like a symphony – it sounds like tuning up.
Applied to an agency's tech stack, that means the quality of any individual AI tool says little about the performance of the whole. What matters is whether data, tools and people talk to one another. Tools alone can't compose a masterpiece; harmony is born from their unison.
That is why the problem usually lies not in the technology but in the architecture. People buy the instruments and forget the score.
What is MCP – and why does it matter to agencies?
MCP stands for Model Context Protocol. It is an open standard that lets AI assistants such as Claude access external data sources and tools. Think of it as a universal socket: instead of building a separate integration or dashboard for every data source, the assistant talks to all of them through a single interface.
For agencies, that changes the logic. Until now, integration meant exporting data, loading it into a tool and copying the result back out. With MCP, the data stays where it is, and the team asks its questions where it already works. ‘Which messages has the competition been running since the start of October?’ becomes a question in the chat, not a ticket to the insight team.
That is exactly how Defy MCP works: teams query competitor, audience and conversation data directly in Claude – like asking a colleague, with no new dashboard to learn. Tinuiti opening up its measurement data to AI agents the same way shows the direction of travel: data is no longer locked away in dashboards but made queryable.
Tool sprawl or integrated stack: the comparison
| Dimension | Tool sprawl | Integrated stack with MCP |
|---|---|---|
| Data access | Exports, copy and paste, screenshots | The assistant queries the source directly |
| Learning curve | One dashboard per tool | One interface, questions in natural language |
| Traceability | Results with no visible origin | Answers can be traced back to the source |
| Governance | Personal accounts, unclear approvals | Central permissions and clear sign-off rules |
| Team knowledge | Lives with individuals | Reusable prompts and workflows |
| Cost | Overlapping licences | Fewer tools, used more deeply |
How do you integrate AI tools in six steps?
- Take an inventory. Who uses which AI tools, for what, with which data? Ask explicitly about personal accounts – without blame, or you won't get an honest answer.
- Pick three time sinks. Typical candidates are competitor audits, monthly reports and pitch research. Record how long they take today.
- Choose one interface. Decide where your team works with AI. Every additional interface needs a good reason.
- Connect data rather than copying it. Check which data sources can be queried directly via MCP or another interface. Anything that only works by export goes on the list for later.
- Set the rules before roll-out. Decide which outputs a human checks before they go to the client, who signs off spend and which client data may go into which systems. Your data protection officer belongs at that table.
- Pilot, measure, clear out. Test for six weeks with one client, compare against the times from step 2 – and cancel the tools nobody misses.
What can independent agencies learn from the holding companies?
What is striking about Digiday's reporting is less the technology than the tone. The agency groups are selling their agents on control: according to Digiday, WPP Media requires explicit controls, approval thresholds and accountable human oversight for financial and activation decisions. At Dentsu, the line is that everything still requires human oversight.
That is a lesson for independents. Clients hesitate less because of the technology than because they fear losing control. If you can explain who checks which output, where a claim comes from and where the data sits, you earn trust – whatever your size.
The difference: holding companies build their stacks with big development budgets. Independent agencies have to combine more cleverly. A platform that brings data, querying and automation together – such as the Defy platform – is one way to reach a comparable level without an in-house data science team.
What should you ask every AI vendor?
- Can your data be queried via MCP or another open interface – or only in your own dashboard?
- Can I trace every claim the system makes back to its source?
- Where is the data stored, who has access, and what happens to our prompts?
- Which workflow does the tool actually replace – and which existing tool can I cancel as a result?
If you don't get an answer to that last question, you are not buying a tool – just another instrument without a score.
