Many teams set their 2027 budgets in the last weeks of 2026. Many AI plans still list tools: a chat seat for each person and a few pilots. In our view, the question changes in 2027. A team that wants to become an AI native company will ask which jobs its agents own, who approves what, and what each agent costs.
This post is a plan for that year. It has five steps. Each step names one decision, how to make it in d5s, the AI-native workspace for teams and their agents, and how you know that the step is done. For what we learned this year, read 4 lessons from teams running AI agents in 2026.
The plan at a glance
- Step 1: Pick repeat jobs for shared agents. Start with work that more than one person does every week.
- Step 2: Set Allow, Ask or Deny for each action. In d5s, reading starts on Allow, and sends and changes start on Ask.
- Step 3: Choose a model per agent. Match the model to the job, and keep a model change to one setting.
- Step 4: Put agents where your team works. Slack, Microsoft Teams, email and calls, not one more app to open.
- Step 5: Measure cost per agent. Give each agent a daily budget, then review its spend, tokens and runs.
- 1Pick repeat jobs for shared agentsEach job has one agent, one owner and one place for results.
- 2Set Allow, Ask or Deny for each actionEach action has a setting that a named person chose.
- 3Choose a model per agentEach agent runs a model that fits its job.
- 4Put agents where your team worksPeople ask the agent in Slack, Teams, email or a call.
- 5Measure cost per agentEach agent has a daily budget and a monthly review.
Step 1: Pick repeat jobs for shared agents
Give your first agents jobs that repeat every week and that more than one person needs. A shared agent builds its context once for the whole team, so each run starts with more of it.
Make a list of the work that comes back each week. Examples are the Monday pipeline review, the month-end close checks, support triage and meeting preparation. For each job, write who needs the result and where it must go. Then pick two or three jobs to start.
Do each job once by hand before you give it to an agent. Ask a team member to do the work in a d5s chat. Then review the sources, the tool activity, the approvals and the usage. When the result is good, turn the pattern into an agent.
In d5s, an agent is a lasting member of the team, not a single chat. It has:
- standing instructions that apply to every run;
- its own memory, files and skills;
- one continuing conversation for direct and scheduled work;
- schedules, and event triggers such as a merged GitHub pull request or a Grafana alert.
Skills spread across the team. When an agent writes a skill in chat and you confirm it, d5s adds that skill to every agent in the workspace.
You can start from a ready-made agent or from a recipe such as the month-end close review.
Done when: each job has one agent, one named owner, and one place where its results go.
Step 2: Set human-in-the-loop approvals for each action
Decide for each action what an agent may do alone. In d5s, reading starts on Allow and sends and changes start on Ask, so a person approves each one until the owner changes the setting.
A rule in a prompt asks the model to wait. A rule in the runtime makes it wait, whatever model runs the agent. So write the rules as settings, not as instructions.
Three settings cover most actions:
- Allow. The agent acts alone. Use it for reads and for low-risk work that you check later.
- Ask. The agent stops and shows an approval card. A person approves or rejects the action.
- Deny. The agent cannot take the action.
In d5s, you set these on the Connectors page under Tool permissions, for each connection and each tool. A connection alone does not give permission. Read Connection is not permission in the docs.
- Read mailDefaultAllow
- Save a draft replyChanged by the ownerAllow
- Send an emailDefaultAsk
- Create a calendar eventDefaultAsk
- Move mail to trashChanged by the ownerDeny
This list is how you govern your agents. We expect more customers and auditors to ask for it in 2027. Keep it short, and keep one owner for it. See how d5s handles approvals and data security.
Done when: each agent has a list of its actions, and each action has a setting that a named person chose.
Step 3: Choose a model per agent
Pick a model for each job, not one model for the company. New models will ship all through 2027, so make a model change one setting, not a new project.
A light, fast model suits frequent, simple work such as triage. A stronger model suits careful work where a mistake costs more, such as a contract check. Stronger models use credits faster, so the choice is also a cost decision.
In d5s, each agent has its own model. Open Manage agent → Profile, then choose Light, Medium or High in the Model menu. To pick a model family, open Advanced. The catalog includes models from OpenAI, Anthropic, Google, Mistral, xAI, DeepSeek and more.
A saved model family uses its newest available version for new turns. Work that already runs keeps its exact version. When you change the model, the agent keeps its instructions, memory, skills and tool permissions.
Before you move an important agent to a new model, test the new model on the same real tasks. Our test of open-source LLMs for AI agents shows one way to do it.
Two more facts for your data review. Team and Enterprise plans use zero data retention on the primary model routes that d5s provides. On Enterprise, d5s can also use your own model provider keys.
Done when: each agent runs a model that fits its job, and you know which test you run before the next change.
Step 4: Put agents where your team works
Bring agents into the tools that your team opens every day. People ask an agent in Slack, Microsoft Teams, email or a call. One more app to open is easy to forget.
A d5s agent can work in:
- Slack and Microsoft Teams channels;
- Telegram and Discord, both in beta;
- email: where available, each new agent gets its own d5s email address, so people can send it work;
- calls: an agent can join Google Meet, Microsoft Teams and Zoom calls, write live notes, and report the outcome in its conversation.
The same agent answers in each place, with the same instructions and the same approval settings. Choose the places from Step 1: put each agent where its results are needed. Read more about channels, agents in email and agents in meetings.
Done when: each agent is in the channel where its results are needed, and the team knows how to ask it.
Step 5: Measure cost per agent
Track spend for each agent, not one model bill. Then you see which jobs pay back, and you can tune the busiest agents first.
Give each agent a daily budget before it works alone. When the agent reaches its budget, it starts no more work that day. The agent is not paused for good.
In d5s, organization owners and admins open Settings → Usage. Agent stats shows the spend, tokens and runs of each agent over time, and includes the runs of its sub-agents. In Activity, open a run to see its cost by model. When the source is still available, Inspect run opens the run trace with each model call.
Review the numbers once a month, with the owner of each agent. Ask two questions:
- Which agents use the most?
- Does their result justify the cost?
Then make one change at a time. Move routine work to a lighter model, or keep the stronger model where errors cost more.
Done when: each agent has a daily budget and an owner who reviews its Agent stats each month.
How to start as an AI native company
- Start from a ready-made agent. Browse all ready-made agents, then change the instructions to fit your team.
- Connect your tools. Browse all connectors, and set the tool permissions for each one.
- Plan the rollout. Read the admin quickstart in the docs before you add more agents.
Questions and answers
What should an AI strategy for 2027 include?
Five decisions: which repeat jobs agents own, what each agent may do alone, which model each agent runs, where the team meets each agent, and what each agent costs. Write one owner next to each decision.
How many AI agents should a team start with?
Few. Start with two or three jobs that repeat every week and that more than one person needs. Do each job once by hand in a chat, then give it to an agent. Add agents when the first ones pay back.
Should AI agents act without approval?
Sometimes. Good AI agent governance lets agents read and do low-risk work alone. Set sends and changes to Ask, and set actions you never want to Deny. In d5s, reading starts on Allow, and sends and changes start on Ask.
Which AI model should our agents use in 2027?
More than one. Use a light model for frequent, simple work and a stronger model where a mistake costs more. In d5s, each agent has its own model, from OpenAI, Anthropic, Google, Mistral, xAI, DeepSeek and more.
How do you control the cost of AI agents?
Per agent. Give each agent a daily budget, then review its spend, tokens and runs in Agent stats each month. Tune the agents that use the most first, for example with a lighter model for routine work.




