Skip to content
Français
Account: not signed in, sync available My learning
Select theme
Français

Recommend changes: prioritise an action plan

By updated

Improve this page

Recommending changes means turning findings into verifiable actions, then ordering them by the risk avoided, the effort required and what the team can absorb. An architecture review that stops at the findings leaves the decision-maker facing twenty problems with no idea which to start with. This lesson is for the Architect track. You start from findings that are already established, write the recommendations, rank them on three criteria (severity, likelihood, effort), separate the quick wins from the structural projects, then build the plan. This is the skill that the assessment at the end of this module scores.

What you will learn

  • Write a recommendation that describes a result and its verification criterion, never a product.
  • Rank recommendations by severity, likelihood and effort, with scales written in advance.
  • Tell a quick win from a structural project, and split an action that is too big.
  • Order a plan taking dependencies, deadlines and team capacity into account.
  • Recognise the most frequent plan mistakes: a list with no ranking, a maximalist recommendation, a risk accepted with no signature.

Prerequisites: the lesson Well-Architected Frameworks, which teaches review by axis, the score backed by evidence and the principle "an accepted risk is signed"; the lesson Design a target architecture, which teaches you to quantify requirements and record a decision in an ADR (Architecture Decision Record, the record that keeps the context, the choice and its consequences); and reading an architecture (flows, state, dependencies, single points of failure), taught by the Services Architecture course (in French). The notions of RTO and RPO come from the lesson Cloud resilience (in French). This page assumes the findings have already been found: it does not teach you to look for them, it teaches what you do with them.

What is a verifiable recommendation?

A verifiable recommendation describes the result to achieve, how to establish that it has been achieved, and what it costs. "Strengthen the security of the database" is not a recommendation: nobody can say one day whether it is done. "No staging machine can open a connection to the production database, as shown by a refused connection test" is one. It can be verified with any provider, because it names a property and not a product.

Six fields are enough to make a recommendation complete. The action says what changes. The verification criterion says how to establish it, by a command, a test or a measurement. The expected effect links the action to the finding it addresses. The effort is expressed in person-days (one day of work by one person), and the cost in euros per month when the action creates or saves any. The residual risk says what remains true once the action is done, because no action brings a risk down to zero: every residual risk must have a named owner who accepts it, with the reason for that choice.

A simple test shows whether a recommendation is finished: the CTO asks "what do I do on Monday morning, how much does it cost, and how will I know it works?". If the text does not answer these three questions, it is still a finding reworded. The example in the lesson writes two recommendations on this model.

How do you rank: severity, likelihood, effort?

Three criteria are enough to order: the severity of what happens if the risk materialises, its likelihood, and the effort to address it. Severity and likelihood multiply into a risk level from 1 to 9. Effort is not added to it: it is then used to break ties and to separate what you do right away from what you plan. Writing the scales before scoring avoids adjusting them to land on the order you already had in mind.

ScoreSeverity (if the risk materialises)Likelihood (over the next twelve months)
1nuisance with no effect on a commitment or on datanot expected within the year
2commitment degraded, or recoverable datamay happen within the year
3contractual commitment broken, data lost or exposedalready happens, or several times a year

These scales are a working proposal, not a standard, and the same goes for the severity times likelihood product: it is a convention of this lesson, common in risk matrices but not required by any standard. An organisation with other stakes (health, finance) will write them differently. What matters is that they are written, shared and applied to all findings. A likelihood score rests, like a review score, on a basis: an incident that already happened, a measurement, or an assumption declared as such.

The risk level does not say everything. A severity of 3 with a low likelihood gives a level of 3, equal to a frequent nuisance, whereas the first breaks a contract. The rule of thumb: every severity 3 stays in the plan, with a date, even at low likelihood. The risk level sorts, it removes nothing.

Quick win or structural project?

A quick win takes two person-days at most, depends on no other action, and can be undone if needed; everything else is a structural project. This boundary is a convention of this lesson, to adapt to the size of the team, but it must be written down. The two categories are not handled the same way. Quick wins are launched the same month, with no committee: they reduce a real risk for almost nothing and show management that the review has an effect.

A structural project changes the shape of the architecture: a second zone, a build pipeline, a component replacement. It is planned, split into verifiable steps, and often depends on another project. It calls for a decision (compare options, see the lesson on architecture decisions) when several paths exist.

A recommendation that is too big gets split. "Isolate staging from production" mixes two actions that do not have the same cost: cutting access (half a day) and providing a replacement dataset (several days). The first is a quick win that removes most of the risk, the second a project that removes the residual risk. Splitting often gives the best ratio of risk avoided to effort, and the coherence of a plan is judged by this fineness.

Worked example: the Salon-Express plan

Salon-Express is an online ticketing service for trade shows: a plan of seven recommendations fits on one page and within the team's capacity. This case is invented for the lesson. There are 40,000 tickets a year, with peaks at the opening of sales. The team has four people. The client contract promises 99.5% availability, an RTO of 2 hours (maximum time to restore the service) and an RPO of 1 hour (maximum data loss). Three constraints frame the plan: the team can devote 6 person-days per month to improvement, the infrastructure budget cannot increase by more than €300 per month, and a client is asking for a compliance audit on personal data in four months.

The review established six findings, which we take as they are. Each is scored with the scales above, the likelihood score resting on the fact given.

No.FindingFact that supports itSeverityLikelihoodRisk
C1Staging uses the production database credentialscontractors access staging; visitor data is readable there326
C2No alert on application errorsthree incidents in two months discovered by customers, 50 minutes on average236
C3Application logs kept with no limit400 GB, +15% per month, €85 per month133
C4Machines built by handthree operating system versions across five machines, a reinstall takes about a day224
C5Application and database in a single zonethe database restore, measured during a test, took 3 h 10 against an RTO of 2 h313
C6The message queue component goes out of support in 14 monthsvendor end-of-support announcement224

Read C5 carefully: its risk level is 3, like that of a nuisance on application logs, but the contractual commitment is already broken by the test measurement. The rule "every severity 3 stays in the plan" protects it from being ranked too low. The seven recommendations that follow address these six findings; C1 produces two, because it is split.

Order of deliveryRecommendationFindingEffortTypeDepends on
1R1. Cut staging's access to the production databaseC10.5 dquick winnothing
2R2. Alert on the error rate and latencyC22 dquick winnothing
3R3. Limit the retention of application logsC30.5 dquick winnothing
4R4. Build the machines through a pipelineC410 dprojectnothing
5R5. Give staging an anonymised datasetC18 dprojectR1
6R6. Spread the application and database across two zonesC515 dprojectR4
7R7. Replace the message queue componentC620 dprojectnothing, deadline at 14 months

R3 only concerns application logs. The audit log, which traces who did what on the account, keeps the duration set by legal obligations and by the client. The site's reference framework cited in the lesson IAM: manage identities and permissions asks for at least six months. This is not a universal rule: the duration depends on legal obligations and on how long it takes to detect an intrusion. The compliance audit expected by Salon-Express's client will need it: shortening its retention to save a few euros would create a new finding.

Here are two of these recommendations written out in full, one of each type, with the six fields.

FieldR1, quick winR6, structural project
ActionRemove the credentials and the network route from staging to the production databasePlace one application machine per zone and a database replica in the second zone
Verification criterionFrom a staging machine, a connection to the production database is refused; the staging configuration no longer contains any production credentialA cut of the main zone is simulated: the service is restored in under 2 hours, stopwatch in hand, with data loss under 1 hour
Expected effectNo visitor data readable from staging any more (addresses C1)The 2-hour RTO commitment is met and measured (addresses C5)
Effort and cost0.5 person-day, €0 per month15 person-days, about +€250 per month (estimate to confirm against the provider's price list)
Residual riskStaging no longer has realistic data: tests suffer until R5A failure that hits both zones or the whole provider remains possible; a second cut test is scheduled every six months
Depends onnothingR4: without reproducible machines, creating the second zone is manual and slow

You need to tell two orders apart. The priority order follows the risk level, from highest to lowest, and at equal level the lowest effort goes first. The order of delivery in the table above is not this simple sort: three forces correct it. First, quick wins go first: R1, R2 and R3 cost three person-days in all, half a month of capacity. Then a dependency imposes the order: R6 addresses the commitment that is already broken, but cannot start before R4, so R4 goes ahead of it. Finally a deadline puts R5 before R6: the audit falls in four months and R5 must be finished before. This choice leaves the RTO commitment broken until the end of R6, in month 6: this risk has to be accepted in writing for the duration of the project. R7, level 4 but with no immediate urgency, is scheduled early enough to finish before the end of support.

Does the plan fit the team's capacity?

A plan is only credible if the sum of the efforts fits within the available capacity. The seven actions represent 56 person-days (0.5 + 2 + 0.5 + 10 + 8 + 15 + 20). At 6 days per month, that makes a little over nine months, so a ten-month plan. The budget is checked the same way: R2 adds about €20 per month, R6 about €250, R3 saves about €60 (estimates to confirm), a balance close to +€210 per month, under the ceiling of €300.

The schedule follows from the dependencies and deadlines: an action only starts once its dependencies are finished, and the load in any period never exceeds the team's capacity. In month 1, the three quick wins (3 days) and the start of R4 (3 days). In months 2 and 3, the end of R4 (7 days) then the start of R5. In month 4, the end of R5 before the audit, then the start of R6. In months 5 and 6, the end of R6. From month 7, R7, which takes twenty days and ends in month 10, several months before the end of support at 14 months. This margin is deliberate: a replacement project often slips.

Each line of the plan stays revisable. An incident, a departure from the team or a client changing their contract moves the order. A useful plan therefore also says under what condition it is reopened, like an ADR, and names the person who arbitrates when capacity is short. A risk you choose not to address is accepted in writing, with a name and a review date.

Which mistakes weaken a plan of recommendations?

Six mistakes come up often, and none depends on the subject. The first is the list with no ranking: twenty findings lined up help nobody decide. The second is the maximalist recommendation, for example "rebuild everything across several regions", which ignores the €300 budget and the 6 days per month. The third is the product in place of the criterion: citing a service without saying which property you are after, which makes the recommendation non-transferable.

The fourth is the plan beyond capacity: seven actions of 56 days promised over three months will not be delivered, and the team will stop believing the plan. The fifth is forgetting dependencies: planning the second zone before the reproducible machines forces you to do everything twice. The sixth is the risk accepted with no signature: a risk set aside with no name or review date is not accepted, it is forgotten. A subtler flaw is added to these: ranking by age or by volume means handling the oldest finding or the most numerous one, not the one that exposes the most.

To go further, the page on security debt shows the same reasoning of prioritisation by risk applied to vulnerabilities. The effort and value matrix of the lesson on migration strategies (in French) applies effort as a second axis for a different decision: the principle is similar, the criteria differ.

How do you check that you can do it alone?

Redo the plan with a capacity of 3 person-days per month: if you can say what changes and who arbitrates, you have mastered the skill. Keep the same seven recommendations and the same dependencies. Ask yourself three questions. Do the seven actions still fit within the audit deadline of four months and the end-of-support deadline of fourteen months? Which recommendation goes ahead of R4 to meet the audit? Which risk must be accepted, and by whom is it signed?

Check your answer: at 3 days per month, the 56 days take nearly 19 months, which exceeds R7's 14-month deadline. The plan does not hold as it stands. The audit does: R1, R2, R3 and R5 take 11 days, against 12 available in four months, provided R5 goes ahead of R4. For the end of support, you have to arbitrate: ask for capacity, push R6 back by accepting in writing an unmet RTO for longer, or split R7. A good answer names the accepted risk, its review date and the person who signs it. It does not just shift the dates without saying anything about what remains exposed.

Put it into practice

The lab gives an architecture review of three findings (publicly accessible database, short retention of audit logs, untested regional recovery) with their efforts and the team's capacity per sprint. The lab provides its own scales, from 1 to 5, different from those of this page: apply its own. You write a plan.json file: one recommendation per finding, the priority order, the classification as quick win or structural project, an owner who accepts each residual risk, then the schedule by sprint. The validation recomputes everything from the review data: risk scores, order, classification according to the rule, presence of the owner and their reason, load of each sprint and respect of dependencies. It checks no real cloud environment. Here, each finding gets a single recommendation: the splitting seen above applies to your own plans.

Key takeaways

  • A recommendation describes a result and its verification criterion, with expected effect, effort in person-days, cost and residual risk; it does not name a product.
  • The risk level is severity times likelihood, on scales written before scoring, each score resting on a fact or a declared assumption.
  • Every severity 3 stays in the plan, even at low likelihood: the risk level sorts, it removes nothing.
  • A quick win costs two person-days at most, depends on nothing and can be undone if needed; a structural project is planned, split and often calls for an architecture decision.
  • Splitting a recommendation that is too big often gives the best ratio between the risk removed and the effort.
  • The order of the plan combines risk, dependencies, deadlines and capacity, and a risk left untreated is accepted in writing, with a person who signs.

Knowledge check

Check that you have grasped the essentials of this guide. The questions cover only what has just been explained here.

Knowledge check

Check what you have learned with this interactive quiz

6 questions
6 min
70% to pass

How it works

  • The timer starts when you click Start
  • Multiple choice, true/false and short answer questions
  • You can move between questions
  • Detailed results are shown at the end

Starts the quiz and the timer